Food Chemistry 159 (2014) 143–150
Contents lists available at ScienceDirect
Food Chemistry journal homepage: www.elsevier.com/locate/foodchem
Honey characterization using computer vision system and artificial neural networks Sahameh Shafiee a, Saeid Minaei a,⇑, Nasrollah Moghaddam-Charkari b, Mohsen Barzegar c a
Department of Agricultural Machinery Engineering, Tarbiat Modares University, Tehran, Iran Department of Electrical and Computer Engineering, Tarbiat Modares University, Tehran, Iran c Department of Food Science and Technology , Tarbiat Modares University, Tehran, Iran b
a r t i c l e
i n f o
Article history: Received 18 November 2013 Received in revised form 22 February 2014 Accepted 24 February 2014 Available online 5 March 2014 Keywords: Computer vision system Colour Honey Antioxidant Ash content
a b s t r a c t This paper reports the development of a computer vision system (CVS) for non-destructive characterization of honey based on colour and its correlated chemical attributes including ash content (AC), antioxidant activity (AA), and total phenolic content (TPC). Artificial neural network (ANN) models were applied to transform RGB values of images to CIE L⁄a⁄b⁄ colourimetric measurements and to predict AC, TPC and AA from colour features of images. The developed ANN models were able to convert RGB values to CIE L⁄a⁄b⁄ colourimetric parameters with low generalization error of 1.01 ± 0.99. In addition, the developed models for prediction of AC, TPC and AA showed high performance based on colour parameters of honey images, as the R2 values for prediction were 0.99, 0.98, and 0.87, for AC, AA and TPC, respectively. The experimental results show the effectiveness and possibility of applying CVS for non-destructive honey characterization by the industry. Ó 2014 Elsevier Ltd. All rights reserved.
1. Introduction Since the earliest times, honey has been used as food, food additive, food preservative and medicine. It is the natural sweet substance produced from the collected nectar of blossoms and exudates of trees or plants by Apis mellifra bees (Alvarez-Suarez, Tulipani, Romandini, Vidal, & Battino, 2009). Honey contains at least 181 substances (Chow, 2002). It is composed of fructose (38%), glucose (31%), minerals, proteins, free amino acids, enzymes and vitamins (Pérez, 2002; Terrab et al., 2003). Also, a wide range of minor ingredients are present in honey, many of which are known to have antioxidant properties, such as phenolic acids (Dimitrova, Gevrenova, & Anklam, 2007; Martos et al., 2000; Tomas-Barberán, Martos, Ferreres, Radovic, & Anklam, 2001). Honey is rich in natural antioxidants and minerals. Thus, it takes a more active role to make a contribution to human health and nutrition. Honey has a wide range of colours from water white to dark amber or dark. Honey colour depends on various factors, mineral content being an important one. Light-coloured honeys usually have low ash content. On the contrary, dark-coloured honeys generally have higher ash contents (Al et al., 2009; Gomes, Dias, Moreira, Rodrigues, & Estevinho, 2010). The evidence of the biological actions of honey can be ascribed to its polyphenolic contents which, in turn, are ⇑ Corresponding author. Tel.: +98 2148292466; fax: +98 2148292200. E-mail address: [email protected] (S. Minaei). http://dx.doi.org/10.1016/j.foodchem.2014.02.136 0308-8146/Ó 2014 Elsevier Ltd. All rights reserved.
usually associated with its antioxidant and anti-inflammatory actions, as well as its cardiovascular, antiproliferative and antimicrobial benefits (Alvarez-suarez, Giampieri, & Battino, 2013). Some research groups have shown that the total phenol and flavonoids contents, and antioxidant activity of honey are greatly dependent on colour (Alvarez-Suarez et al., 2010; Bertoncelj, Doberšek, Jamnik, & Golob, 2007; Estevinho, Pereira, Moreira, Dias, & Pereira, 2008; Ferreira, Aires, Barreira, & Estevinho, 2009). Therefore, colour is an important feature in quality control of honey. The quality parameters (i.e., antioxidants) of honey are normally measured using conventional analytical techniques (Alvarez-Suarez et al., 2009). However, applying these techniques in honey industry has some disadvantages such as their destructive nature, implementation expense, and a time requirement. For food authentication and honey characterization in particular, introduction of alternative methods to conventional procedures aiming at objective measurement of honey in a consistent and cost effective manner is of great importance for the honey industry (Shafiee, Minaei, MoghaddamCharkari, Ghasemi-Varnamkhasti, & Barzegar, 2013). In this regard, several studies have been carried out to develop nondestructive techniques using a calibration data set obtained by analytical methods for various agricultural products (Brosnan & Sun, 2004; Butz, Hofmann, & Tauscher, 2005; Pace et al., 2013; Ruiz-Altisent et al., 2010). Among these approaches, computer vision can be noted for having good capability of food colour assessment as well as food characterization for those properties associated with
144
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
colour. Many research groups have used computer vision for analysis of food colour. The first step in the development of a computer vision technique for food quality assessment based on colour is colour calibration and characterization. Briones and Aguilera (2005) employed image analysis techniques to analyse colour variations during blooming development in chocolate. They concluded that computer vision systems have the capability to quantify overall changes as well as particular features over the whole chocolate surface. Thus, they introduced their vision based approach for customization and standardization of quality assessment of chocolate. Zheng et al. (2011) proposed a method to predict the changes of anthocyanins, ascorbic acid, total phenols, flavonoids, and antioxidant activity of red bayberry juice using CVS and ANN combination. In a recent study, CVS has been utilized to predict antioxidant activity and total phenols in pigmented carrots (Pace et al., 2013). It was demonstrated that CVS can act as an effective colour assessment tool and act as a proper predictor for antioxidant properties of carrot pigments. On the other hand, there is no report that employs CVS for predicting honey antioxidant activity, total phenolic content, and ash content. The objective of this research was to study the relationships between TPC, AA, AC, and honey colour. Moreover, we aim to evaluate the effectiveness of CVS–ANN combination for honey colour assessment and prediction of its TPC, AA, and AC. 2. Materials and methods 2.1. Honey samples and chemicals One hundred twenty nine honey samples of various floral origins and colours were analysed. These included: Loco (Astragalus bisulcatus), Opoponax-Tree, Alfalfa (Medicago sativa), Barberry (Berberis vulgaris), Thyme (Thymus vulgaris), Argentine thistle (Eryngium billardieri), and Dill (Anethum graveolens dhi). Samples were obtained directly from bee keepers of several different provinces of Iran. All chemicals and solvents used for analysis were analytical grade with the highest purity available. Folin–Ciocalteu’s phenol reagent, gallic acid, sodium carbonate anhydrous, sodium acetate trihydrate, ferrous ammonium sulphate, and ferric chloride hexahydrate were purchased from Merck (Merck KGaA, Darmstadt, Germany). Furthermore, 2,4,6-tripyridyl-s-triazine (TPTZ) was purchased from Sigma–Aldrich Chemie GmbH (Steinheim, Germany). 2.2. Colour measurement using Hunter lab and determination of ash content (AC) Before colour analysis, samples were heated at 50 °C for 10 min to dissolve sugar crystals. Then, they were paved and pureed into 5 cm Petri dishes at 1 cm height. Colour characteristics were measured using the CIE L⁄a⁄b⁄ colourimetric method. Honey colour values were determined using a commercial colourimeter (Hunter Lab, color Flex, USA). Percent ash content was determined for all honey samples based on AOAC (1990). 2.3. Total phenolic content (TPC) The Folin–Ciocalteu method was utilized to determine the total Phenolic content of the samples (Alvarez-Suarez et al., 2010). Honey solution in distilled water was prepared (10% w/v) and 0.5 ml of this solution was poured into a test tube. Then, 2.5 ml of Folin–Ciocalteu 0.2 N was added and well mixed. After 5 min, 2 ml of 0.7 M sodium carbonate was added. This solution was incubated in dark room at 25 °C for 2 h. Absorbance of the reaction mixture was measured at 760 nm against the sugar analogue using
spectrophotometer (Jenway, model 6505). Gallic acid was used as the standard to produce the calibration curve (50–250 mg/L). The total Phenolic content was expressed in mg of Gallic acid equivalents (mg GAE/kg of honey). 2.4. Quantification of antioxidant activity (AA) Several methods for determining the anti-oxidative activity in honey have been employed. The techniques to evaluate antioxidant capacity are based on colourimetric assays such as DPPH, FRAP, TEAC (ABTS) and microplate fluorescence reader like ORAC assay (Alvarez-Suarez et al., 2009). Antioxidants stabilize free radicals by donating electrons and the antioxidant capacity provided by this mechanism is determined by measuring the reduction capacity of metal ions such as ferric ion (Fe3+): FRAP (Niki, 2011). In this study, FRAP assay has been used for determination of honey antioxidant capacity (ferric reducing/antioxidant power). It is a simple direct test that widely used for assessment of antioxidant activity of various materials including honey (Bertoncelj et al., 2007). 2.4.1. The ferric reducing antioxidant power assay Benzie and Strain (1996) procedure with minor modification was used for FRAP assay. The principle of this method is based on the reduction of a ferric 2,4,6-tripyridyl-s-triazine complex (Fe3+-TPTZ) to its ferrous, coloured form (Fe2+-TPTZ) in the presence of antioxidants. The FRAP reagent contained 2.5 ml of a 10 mM TPTZ (2,4,6-tripyridyl-s-triazine) solution in 40 mM HCl, 2.5 ml of 20 mM FeCl3 and 25 ml of 0.3 M acetate buffer, pH 3.6. It was prepared daily and warmed to 37 °C. Aliquots of 400 ll of sample were mixed with 3.6 ml of FRAP reagent and the absorbance of the reaction mixture was measured at 593 nm after incubation at 37 °C for 10 min against the sugar analogue. Ammonium ferrous sulphate (100–1000 lM) was used for the calibration curves and the results were expressed as lmoles of ammonium ferrous sulphate per 100 g of honey (lmol Fe (II)/100 g of honey). All tests were carried out in triplicate and their averages were reported. PASW statistics software was utilized for statistical analysis. 2.5. Computer vision system (CVS) The computer vision system used in this work consists of four major components: dark chamber and lighting system, digital camera, computer hardware and software. A 30 30 45 cm3 wooden box was prepared to act as the housing. The bottom part of this box (15 cm) was used to form the illuminating system. Furthermore, the upper 30 cm was used to form the dark chamber. A plastic light diffuser sheet which is used to hold the sample separates these two sections from each other. Back-lighting was used for illumination of honey samples to diminish the negative effect of reflectance and to create homogenous light intensity on the honey sample surface. The back-lighting system consisted of a 30 30 15 cm3 box with its internal surfaces covered using a reflective sheet for more light scattering in the lighting chamber. LED lamps were placed horizontally in the chamber. A plastic light diffuser was placed on top of the chamber. A Canon 550 D kiss x4 camera was mounted vertically above the back lighting chamber at a distance of 20 cm. The camera was connected to the USB port of a Laptop (Dell, Inspiron 4030, China) provided with a remote capture software (Zoom Browser Ex. Version: 6.7) to view and to capture images from the camera setting controller. The internal surfaces of the wooden box were covered with dark sheets to prevent reflection and external light. The Honey sample was located against the center of the background (Fig. 1a). A standard grey card (10 12 cm2) with 18% reflectance (Kodak, USA) was used to set
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
145
Fig. 1. The scheme designed for prediction of the chemical attributes of honey using computer vision system and artificial neural networks.
the white balance of the CCD camera. Various different camera settings (combination of aperture, shutter speed and ISO) was tried to arrive at the average L⁄ pixel value and standard deviation of 48.1 ± 0.2 for the grey card. In order to achieve the illumination system stabilization, it was switched on 30 min prior to image acquisition. Images were captured using the following camera settings: exposure mode: manual, shutter speed: 0.3, lens aperture setting: 5.6, image resolution: 5184 3486 pixels, picture style: standard, and colour space: RGB. Camera was manually focused and the zoom was adjusted. Captured images were stored in bitmap format. All images were captured in triplicate.
pixels) was extracted from the center of each honey image (Fig. 1c). Bubbles and bee wax were removed from each image during the preprocessing step. Otsu thresholding method was used for separation of particles from the background (Fig. 1d). Then, the resulting image was subtracted from the original image. Finally, voids in the resulting image were filled with their neighbourhood intensities (Fig. 1e). This is plausible, because of the homogenous nature of the honey surface and close resemblance between the image intensities of neighbouring pixels. The RGB values for each colour patches were calculated by averaging RGB values of all pixels of processed cropped image.
2.6. Image processing
2.7. Colour transformation
The image preprocessing steps are presented in Fig. 1. After image acquisition and storage (Fig. 1b), crop image (1751 1275
Various colour spaces have been introduced within the last decade. Each of these colour spaces are formed by a linear or
146
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
non-linear transformation of RGB colour space. Because of their specific characterization, HSV, YIQ, CIE L⁄a⁄b⁄, and YCbCr transformations have been implemented in this study. 2.7.1. YIQ colour space A colour space known as YIQ is used in NTSC1. One of the main advantages of this format is that the intensity component is separated from colour data, such that the same signal can be used for both colour and black and white sets. In NTSC colour space, image data consists of three components: luminance (Y), hue (I), and saturation (Q). The first component represents greyscale information, while the last two components represent chrominance (colour information). To obtain YIQ values, the following transformation matrix is used:
2
3 2 3 2 3 Y 0:299 0:587 0:114 R 6 7 6 7 6 7 4 I 5 ¼ 4 0:596 0:274 0:322 5 4 G 5 Q 0:211 0:523 0:312 B 2.7.2. YCbCr colour space This colour space is widely used in digital video systems. Accordingly, luminance information is stored as a single component (Y), and chrominance information is stored as two colour-difference components (i.e., Cb and Cr). Cb represents the difference between the blue component and a reference value. Cr represents the difference between the red component and a reference value. The YCbCr transformation matrix is given below (Du & Sun, 2005):
2
3 2 3 2 3 Y 0:2989 0:5866 0:1145 R 6 7 6 7 6 7 4 Cb 5 ¼ 4 0:1688 0:3312 0:5000 5 4 G 5 Cr 0:5000 0:4184 0:0816 B 2.7.3. CIE L⁄ a⁄ b⁄ Based on CIE (1931) XYZ, CIE L⁄a⁄b⁄ was developed to measure the colour differences consistently with the perceived colour differences. The L⁄ channel stands for brightness, while a⁄ and b⁄ channels define the content of colour information, in which a⁄ extends from green to red and b⁄ from blue to yellow. It has been found that the Euclidian distance in L⁄a⁄b⁄ space provides better match to the human visual perception of colour. 2.7.4. HSV colour space HSV stands for hue, saturation, and value. It was introduced with the aim of providing a more intuitive colour mixing model. HSV attempts to capture the colour components, the way humans perceive colour. Furthermore, it can be found by using a simple transformation which makes it possible to use in real time applications. 2.8. Artificial neural networks (ANN) Colour calibration is concerned with adjusting the imaging system to a known state producing consistent and reproducible colour values. Characterization is the relationship between image primaries (RGB) and a device-independent colour space, such as the CIE XYZ or CIE LAB (Johnson, 1995; Westland & Ripamonti, 2004). In this sense, one way to use a digital camera as a colourimeter is to find a model, often some type of regression which directly maps RGB values to CIE XYZ or CIE LAB (Valous, Mendoza, Sun, & Allen, 2009). In this study, artificial neural networks technique was utilized for characterization of honey samples, and RGB mapping to CIE L⁄a⁄b⁄. Also, the prediction of honey TPC, AA, and AC were 1
National television systems committee.
done by ANN. A network with a single hidden layer was used to characterize honey colour. In different studies it has been shown that a single hidden layer would be sufficient for ANN to approximate any complex, nonlinear function (Dogan, Demirpence, & Cobaner, 2008). In this study, two feed-forward networks with input, output, and hidden layers were used for colour characterization and TPC, AA, and AC prediction. The Input layer of the network for honey colour measurement consisted of three neurons which corresponded to R, G, and B values extracted from honey sample images. The output layer had one neuron representing L⁄, a⁄, and b⁄ values of honey samples. Moreover, the network for TPC, AA, and AC predictions had 15 neurons in the input layer corresponding to 15 colour features extracted from each honey image (R, G, B, H, S, V, Y, I, Q, Y, Cb, Cr, L, a, and b) and one neuron in the output layer representing of TPC, AA, and AC of honey samples. Backpropagation algorithm was utilized in training of the ANN models. Because of the choice of the back-propagation network, the input data was normalized so that it ranged from zero to one in value before being fed to the ANN. Tangent-sigmoid and log-sigmoid transfer functions were used in the hidden layer and the output layer, respectively. Error minimization was accomplished using the Levenberg–Marquardt (LM) algorithm. A trial-and-error approach was utilized to set the number of neurons in the hidden layer. The Optimum number of nodes in the hidden layer was defined based on the lowest root mean square error value, RMSE. In this regard for any number of nodes in the hidden layer, execution of the ANN algorithm was repeated 10 times and the average RSME was calculated. For comparison of the actual values of CIEL⁄a⁄b⁄measured by the colourimeter and those predicted by the developed ANN model, colour differences were calculated. The average values of CIEL⁄a⁄b⁄ colour differences for validation and test sets was referred to as generalization error. The value of colour difference, DEab (CIE, 1986) was computed using the following equation:
DEab ¼
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 2 ðDL Þ þ ðDa Þ2 þ ðDb Þ
ð3Þ
The required codes were developed using MATLAB 2012 software. A schematic diagram of all stages for prediction of AA, TPC, and AC by the CVS and ANN combination is shown in Fig. 1. 3. Results and discussion 3.1. Honey colour and its relationships with TPC, AA, and AC Colour and chemical attributes of honey samples were determined as shown in Table 1. Colours of the honey samples were noticeably different and varied from white to dark amber. The brightest is Alfalfa honey that has the white colour to extra light amber tones. Moreover, the darkest one is Barberry honey that has the amber to dark amber colour tones. The mean value and standard deviation of L⁄, a⁄, and b⁄ colour parameters of honeys obtained using Hunter lab colourimeter are summarized in Table 1. Alfalfa, Loco, and Opponax samples had the highest average values of lightness indicated by L⁄ parameter (57.30, 56.23 and 54.22, respectively). There is no significant difference between these three types of honey and visual comparison confirms this. The dark amber Barberry honey was found to be the darkest having minimum value of lightness (20.06). Dill and Thyme honey are amber. In a report by González-Miret, Terrab, Hernanz, FernándezReca-males, and Heredia (2005) honeys were classified into two groups based on their lightness value (light honeys where L⁄ > 50) and dark honeys (where L⁄ < 50). In this regard, Alfalfa, Loco, Opponax and Argentine thistle are light honeys while Barberry, Dill and Thyme honeys are categorize as dark. Other colour parameter, a⁄ and b⁄ values, varied from 20.51 to 25.07
147
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150 Table 1 Mean ± STD values of colour parameters (L⁄a⁄b⁄ measured by Hunter Lab colourimeter), TPC, AA, and ash content of honey samples. Type of honey
Number of samples
L⁄
a⁄
b⁄
Ash content (%)
Total phenol content (mggallic kg)
1 2 3 4 5 6 7 Total
24 6 9 27 21 27 15 129
56.23 ± 3.23d 46.47 ± 3.56b 20.06 ± 1.03a 57.30 ± 3.46d 47.37 ± 4.87b 54.22 ± 3.92c,d 51.05 ± 5.98c
3.66 ± 1.38a,b 5.39 ± 1.32a,b 25.07 ± 1.11d 2.51 ± 0.89a 12.57 ± 3.01d 7.68 ± 1.76b,c 9.18 ± 1.57c,d
45.01 ± 4.06b,c 40.40 ± 7.78a,b,c 25.85 ± 1.31a 41.58 ± 6.30b 55.36 ± 4.68d 50.39 ± 7.93c,d 54.47 ± 3.58d
0.08 ± 0.02a,b 0.10 ± 0.03b 0.82 ± 0.01d 0.05 ± 0.02a 0.16 ± 0.02c 0.11 ± 0.06b 0.08 ± 0.01a,b
37.67 ± 5.53a 74.72 ± 7.63c 259.52 ± 34.49d 33.34 ± 5.96a 75.17 ± 5.11c 49.93 ± 10.2a,b 57.26 ± 11.45b,c
acid/
FRAP (lmol Fe (II)/ 100 g) 256.27 ± 16.11b 678.84 ± 41.87f 1383.18 ± 94.2g 204.14 ± 19.9a 558.02 ± 34.04e 325.28 ± 23.73c 431.37 ± 25.6d
Data with common letters have no significant differences (Duncan a < 0.05).
and 25.85 to 55.36 among the seven types of honey. There is a noteworthy difference between some types of honey for a⁄ and b⁄ parameters as shown in Table 1. Antioxidant activity of honey samples assessed by FRAP assay is entirely different (p < 0.05) among the seven types of honey. Barberry honey which is the darkest of all had the highest antioxidant activity and Alfalfa honey being the lightest honey had minimum antioxidant activity. Results showed that the TPC varied among different types of honey as its minimum value was for Alfalfa honey (33.34) which is the lightest. Moreover, the maximum value of TPC was associated with barberry, the darkest. The results confirmed the study done by Alvarez-Suarez et al. (2009) for Cuban honeys indicating that the darker samples includes higher AA and TPC values. Also, it has been found that the darker the honey, the higher the ash content. Thus the ash content of Barberry and Dill honeys are higher than that of the Alfalfa and Loco. This corroborates the results obtained by other researchers. Al et al. (2009) indicated that light-coloured honeys usually have low ash content while dark-coloured honey generally contain higher amount of ash. The relationships between honey colour and TPC, AA, and AC are presented in the correlation matrix below (presenting Pearson correlation coefficients).
2
L
a
Phenol content ash content
6 0:76 6a 6 6 Phenol content 0:78 0:73 6 6 0:70 0:64 0:56 4 ash content FRAP value
0:85 0:61 0:72
3 7 7 7 7 7 7 5
0:77
As it is found, all the relationships are statistically significant (p < 0.01). The highest correlation coefficients found is associated with the relationship between FRAP antioxidant activity and the L⁄ Value of honey colour (0.85), followed by TPC and L⁄ (0.78). These results confirm those obtained by Bertoncelj et al. (2007) for Slovenian honeys. Finally, The lower correlation (0.7) was found for ash content and L⁄ values of honey samples in comparison with FRAP value and total phenol content. The results of previous studies and this research show that honey colour is an important feature for honey characterization. Thus, it is possible to measure honey colour by a proper instrument and characterize honey for different colour-dependent parameters such as total phenol content, antioxidant activity, and ash content.
Table 2 Root mean square error of different number of nodes in the hidden layer for colour and chemical attributes of honey samples (mean ± standard deviation). Number of nodes in the hidden layer
1 2 3 4 5 6 7 8 9 10
RMSE L⁄
a⁄
b⁄
AC
TPC
AA
1.80 ± 0.48 1.83 ± 0.34 1.50 ± 0.11 1.59 ± 0.30 1.50 ± 0.16 1.73 ± 0.37 1.49 ± 0.54 1.91 ± 0.89 1.67 ± 0.34 1.42 ± 0.42
1.57 ± 0.40 1.60 ± 0.48 1.53 ± 0.34 1.65 ± 0.58 1.36 ± 0.29 1.48 ± 0.169 1.57 ± 0.630 1.41 ± 0.193 1.52 ± 0.665 1.38 ± 0.251
1.74 ± 0.14 2.15 ± 0.66 1.80 ± 0.30 1.99 ± 0.68 1.78 ± 0.66 1.90 ± 0.63 1.99 ± 0.93 1.47 ± 0.32 1.55 ± 0.28 1.38 ± 0.23
0.114 ± 0.019 0.109 ± 0.019 0.100 ± 0.015 0.107 ± 0. 01 0.107 ± 0.01 0.102 ± 0.01 0.091 ± 0.02 0.138 ± 0.03 0.080 ± 0.02 0.095 ± 0.01
4.02 ± 0.05 3.53 ± 0.38 3.09 ± 0.34 1.59 ± 0.21 1.56 ± 0.21 3.90 ± 0.41 3.31 ± 0.36 2.52 ± 0.31 1.65 ± 0.26 2.13 ± 0.22
7.19 ± 1.30 6.91 ± 1.63 5.46 ± 0.50 5.68 ± 0.72 5.88 ± 1.52 5.40 ± 0.97 6.62 ± 1.59 5.89 ± 1.77 5.44 ± 1.28 5.39 ± 0.82
Fig. 2. Correlations between the predicted and measured values of L⁄, a⁄ and b⁄ in the training set.
148
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
Fig. 3. Change in the values of DEab for each honey sample, considering the test and validation set of ANN.
3.2. Computer vision system and artificial neural network combination (CVS–ANN) The main objective of this study was to prove the applicability of CVS–ANN combination as a simple and potential tool for colour measurement and prediction of TPC, AA, and AC. In this regard, a proper computer vision system was developed and calibrated. Consistency of illumination system is an important factor, since colour reproducibility is dependent on this factor. A preliminary test was conducted for sensitivity analysis of the lighting system indicated the uniformity of illumination setup over the effective illuminated
area in the center of image. The value of DL⁄ for images captured in 20 min intervals during 2 h was lower than 0.09. This shows the good stability of illumination with time. The Homogeneously illuminated area of the grey card was a 8 12 cm2 rectangle in which the L⁄ value was at a consistent level of 48.1 (data not shown). After camera and lighting system calibrations, performing image processing, and colour transformation, ANN was used for colour characterization and TPC, AA, and AC prediction for honey samples. Network topology is an important factor in the design of ANNs since the topology type has a significant influence on the final network accuracy. The number of neurons in the hidden layer is one of the important factors in designing MLP networks with minimum error and maximum accuracy. In order to select the optimum number of neurons for the hidden layer, training step started with 1 neuron and was increased to 10 neurons. The original data set of 129 honey samples was divided into ten blocks, 70% of each block of data was randomly selected to generate the model. Data validation was performed with 15% of the total data (selected randomly from each block). In addition, the remaining 15% of data were used in the test phase. Table 2 shows the variation of RMSE in L⁄, a⁄, and b⁄ assessment, and the prediction of TPC, AA, and AC with various number of nodes in the hidden layer based on different colour features (15 different colour values for TPC, AA, and AC prediction). As shown in Table 2, the desired number of nodes in the hidden layer are 10, 5, 10, 9, 5, 10, respectively, for L⁄, a⁄, b⁄ assessments and, AC, TPC, and AA predictions.
Fig. 4. Correlations between the predicted and experimental values of AC, AA, and TPC (a, d, h) training set, (b, e, h) validation set, (c, f, i) test set.
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
Correlations between the experimental and predicted values of L⁄, a⁄ and b⁄ for the training set of honey samples are shown in Fig. 2. It can be seen that there is a proper correlations between ANN output and the actual measured values of honey colour parameters with R2 of 0.999, 0.999, and 0.974, and RMSE of 0.23, 0.14 and 1.55 for L⁄, a⁄ and b⁄. The CVS- ANN combination exhibited good prediction for the validation and test sets. Variation of, DEab for different honey samples (38 honey sample) in validation and test sets of ANN are shown in Fig. 3. The mean and standard deviation value for generalization error can be given as 1.01 ± 0.99. The minimum average distinguishable DEab by human eye has been reported to be 2.2 (Brainard, 2003). In this study, the generalization error average and standard deviation are below the threshold value of 2.2 (Fig. 3). This shows the effectiveness of the developed ANN model for honey colour measurement using CVS. Previous studies have shown the efficiency of CVS for food colour measurements. Valous et al., 2009 used colour calibration of a laboratory computer vision system for quality evaluation of pre-sliced hams. Their results showed that CVS can be used as an objective tool to specify colour measurement of ham slices. Thus it can be stated that, a CVS–ANN combination can be used for precise quantitative colour measurement of diverse food materials including honey. Correlations between the predicted and experimental AC, TPC and AA values for the training, validation and test sets are shown in Fig. 4 (a,d,g), (b,e,h) and (c,f,i), respectively. It can be seen that the optimal ANN could predict the AC, AA and TPC values of honey samples on the basis of colour data measured using CVS with reasonable R2 values of 0.99, 0.98, 0.87 and RMSE of 0.08, 1.58, and 5.4, respectively. These results show that colour parameters are good predictors for honey total phenol content, antioxidant activity and ash content. Furthermore CVS–ANN combination is a good tool for nondestructive measurement of honey nutrients. These results confirm those obtained by Zheng et al. (2011) for predicting anthocyanins and antioxidant activity of bayberry juice during storage. They reported that red, green, and blue intensity values are good predictors for anthocyanins and antioxidant activity parameters. Similar results were obtained by Pace et al. (2013) for pigmented carrots. They proposed two multiple regression models to predict AA and TPC of pigmented carrots on the basis of colour data measured using CVS. It was concluded that CVS provided more stable and informative colour evaluation, even for unevenly coloured surfaces compared to colourimeter. Our study results indicate that CVS–ANN can be effectively used for honey colour and nutrient assessment. 4. Conclusion This study was conducted for non-destructive characterization of honey using computer vision system. Since, ash content, antioxidant activity and total phenolic content of honey affect its colour; a computer vision system was used for prediction of these characteristics in honey samples using colour attributes of honey images. The proposed system measures honey colour with low generalization error. Thus, CVS can be used as a precise and fast colourimeter for honey colour assessment. Developed models for nutrition measurement of honey were able to predict AC, AA, and TPC with high regression coefficients. Our results show the accuracy and efficiency of CVS–ANN combination for non-destructive characterization of honey on a laboratory scale. A larger dataset — a large number of honey samples with various floral and geographical origins — can be used for programming the CVS which would be further use for industrial scale applications. This can be extended to other food products whose quality is associated with their colour, shape or texture.
149
Acknowledgemens The authors would like to express their appreciation to Dr. Mohammad Reza Khoshayand of Tehran University of Medical Sciences, Faculty of Pharmacy for providing the required substances for chemical tests. References Al, M. L., Daniel, D., Moise, A., Bobis, O., Laslo, L., & Bogdanov, S. (2009). Physicochemical and bioactive properties of different floral origin honeys from Romania. Food Chemistry, 112(4), 863–867. Alvarez-suarez, J. M., Giampieri, F., & Battino, M. (2013). Honey as a source of dietary antioxidants: Structures, bioavailability and evidence of protective effects against human chronic diseases. Current Medicinal Chemistry, 20(5), 621–638. Alvarez-Suarez, M., Tulipani, S., Díaz, D., Estevez, Y., Romandini, S., Giampieri, F., et al. (2010). Antioxidant and antimicrobial capacity of several monofloral Cuban honeys and their correlation with color, polyphenol content and other chemical compounds. Food and Chemical Toxicology, 48, 2490–2499. Alvarez-Suarez, J. M., Tulipani, S., Romandini, S., Vidal, A., & Battino, M. (2009). Methodological aspects about determination of compounds and in vitro evaluation of antioxidant capacity in the honey: A review. Current Analytical Chemistry, 5, 293–302. AOAC. (1990). Official methods of analysis of the association of analytical (15th ed.). Washington, DC: AOAC. Benzie, I. F., & Strain, J. (1996). The ferric reducing ability of plasma (FRAP) as a measure of ‘‘Antioxidant Power’’: The FRAP assay. Analytical Biochemistry, 239, 70–76. Bertoncelj, J., Doberšek, U., Jamnik, M., & Golob, J. (2007). Evaluation of the phenolic content, antioxidant activity, and color of Slovenian honey. Food Chemistry, 105, 822–828. Brainard, D. H. (2003). Color appearance and color difference specification. In S. K. Shevell (Ed.), The science of color (2nd ed., pp. 191–216). Amsterdam: Elsevier Science Ltd. Briones, V., & Aguilera, J. M. (2005). Image analysis of changes in surface color of chocolate. Food Research International, 38, 87–94. Brosnan, T., & Sun, D. W. (2004). Improving quality inspection of food products by computer vision – A review. Journal of Food Engineering, 61, 3–16. Butz, P., Hofmann, C., & Tauscher, B. (2005). Recent developments in non-invasive techniques for fresh fruit and vegetable internal quality analysis. Journal of Food Science, 70, 131–134. Chow, J. (2002). Probiotics and prebiotics: A brief overview. Journal of Renal Nutrition, 12, 76–86. CIE (1931). International Commission on Illumination. In Proceeding of the 8th session, Cambridge, England, UK. CIE (1986). Colorimetry – Official recommendations of the international commission on illumination. CIE Publication No. 15.2, CIE Central Bureau, Vienna, Austria. Dimitrova, B., Gevrenova, R., & Anklam, E. (2007). Analysis of acids in honeys of different floral origin by solid-phase extraction and high-performance liquid chromatography. Phytochemical Analysis, 18, 24–32. Dogan, A., Demirpence, H., & Cobaner, M. (2008). Prediction of groundwater levels from lake levels and climate data using ANN approach. Water SA, 34(2), 1–10. Du, C., & Sun, D. (2005). Comparison of three methods for classification of pizza topping using different colour spaces transformations. Journal of Food Engineering, 68, 277–287. Estevinho, L., Pereira, A. P., Moreira, L., Dias, L. G., & Pereira, E. (2008). Antioxidant and antimicrobial effects of compounds extracts of Northeast Portugal honey. Food and Chemical Toxicology, 46, 3774–3779. Ferreira, I. C. F. R., Aires, E., Barreira, J. C. M., & Estevinho, L. M. (2009). Antioxidant activity of Portuguese honey samples: Different contributions of the entire honey and extract. Food Chemistry, 114, 1438–1443. Gomes, S., Dias, L. G., Moreira, L. L., Rodrigues, P., & Estevinho, L. (2010). Physicochemical, microbiological and antimicrobial properties of commercial honeys from Portugal. Food and Chemical Toxicology, 48(2), 544–548. González-Miret, M. L., Terrab, A., Hernanz, D., Fernández-Reca-males, M. A., & Heredia, F. J. (2005). Multivariate correlation between color and mineral composition of honeys and by their botanical origin. Journal of Agricultural and Food Chemistry, 53, 2574–2580. Johnson, T. (1995). Methods for characterizing colour scanners and digital cameras. Displays, 16(4), 183–191. Martos, I., Ferreres, F., Yao, L., D’Arcy, B., Caffin, N., & Tomas-Barberan, F. A. (2000). Flavonoids in mono specific Eucalyptus honeys from Australia. Journal of Agricultural and Food Chemistry, 48, 4744–4748. Niki, E. (2011). Antioxidant capacity: Which capacity and how to assess it ? Journal of Berry Research, 1(4), 169–176. Pace, B., Cefola, M., Renna, F., Renna, M., Serio, F., & Attolico, G. (2013). Multiple regression models and computer vision systems to predict antioxidant activity and total phenols in pigmented carrots. Journal of Food Engineering, 117, 74–81. Pérez, R. A. (2002). Analysis of volatiles from Spanish honeys by solid-phase microextraction and gas chromatography–mass spectrometry. Journal of Agricultural and Food Chemistry, 50, 2633–2637.
150
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
Ruiz-Altisent, M., Ruiz-Garcia, L., Moreda, G. P., Lu, R., Hernandez-Sanchez, N., Correa, E. C., et al. (2010). Sensors for product characterization and quality of specialty crops – a review. Computers and Electronics in Agriculture, 74, 176–194. Shafiee, S., Minaei, S., Moghaddam-Charkari, N., Ghasemi-Varnamkhasti, M., & Barzegar, M. (2013). Potential application of machine vision to honey characterization. Trends in Food Science & Technology, 30, 174–177. Terrab, A., Gonzále, M. L., González, A. G., Díez, M. F., Escudero, M. L., & Heredia, F. J. (2003). Characterisation of Moroccan unifloral honeys using multivariate analysis. Europeans Food Research and Technology, 218, 88–95. Tomas-Barberán, F. A., Martos, I., Ferreres, F., Radovic, B. S., & Anklam, E. (2001). HPLC flavonoid profiles as markers for the botanical origin of European unifloral honeys. Journal of the Science of Food and Agriculture, 81, 485–496.
Valous, N. A., Mendoza, F., Sun, D., & Allen, P. (2009). Colour calibration of a laboratory computer vision system for quality evaluation of pre-sliced hams. Meat Science, 81, 132–141. Westland, S., & Ripamonti, C. (2004). Computational colour science using matlab. Chichester: John Wiley and Sons, pp. 111–140. Zheng, H., Jiang, L., Lou, H., Hu, Y., Kong, X., & Lu, H. (2011). Application of artificial neural network (ANN) and partial least-squares regression (PLSR) to predict the changes of anthocyanins, Ascorbic acid, total phenols, flavonoids, and antioxidant activity during storage of red bayberry juice based on fractal analysis and red, green, and blue (RGB) intensity values. Journal of Agricultural and Food Chemistry, 59, 592–600.
Contents lists available at ScienceDirect
Food Chemistry journal homepage: www.elsevier.com/locate/foodchem
Honey characterization using computer vision system and artificial neural networks Sahameh Shafiee a, Saeid Minaei a,⇑, Nasrollah Moghaddam-Charkari b, Mohsen Barzegar c a
Department of Agricultural Machinery Engineering, Tarbiat Modares University, Tehran, Iran Department of Electrical and Computer Engineering, Tarbiat Modares University, Tehran, Iran c Department of Food Science and Technology , Tarbiat Modares University, Tehran, Iran b
a r t i c l e
i n f o
Article history: Received 18 November 2013 Received in revised form 22 February 2014 Accepted 24 February 2014 Available online 5 March 2014 Keywords: Computer vision system Colour Honey Antioxidant Ash content
a b s t r a c t This paper reports the development of a computer vision system (CVS) for non-destructive characterization of honey based on colour and its correlated chemical attributes including ash content (AC), antioxidant activity (AA), and total phenolic content (TPC). Artificial neural network (ANN) models were applied to transform RGB values of images to CIE L⁄a⁄b⁄ colourimetric measurements and to predict AC, TPC and AA from colour features of images. The developed ANN models were able to convert RGB values to CIE L⁄a⁄b⁄ colourimetric parameters with low generalization error of 1.01 ± 0.99. In addition, the developed models for prediction of AC, TPC and AA showed high performance based on colour parameters of honey images, as the R2 values for prediction were 0.99, 0.98, and 0.87, for AC, AA and TPC, respectively. The experimental results show the effectiveness and possibility of applying CVS for non-destructive honey characterization by the industry. Ó 2014 Elsevier Ltd. All rights reserved.
1. Introduction Since the earliest times, honey has been used as food, food additive, food preservative and medicine. It is the natural sweet substance produced from the collected nectar of blossoms and exudates of trees or plants by Apis mellifra bees (Alvarez-Suarez, Tulipani, Romandini, Vidal, & Battino, 2009). Honey contains at least 181 substances (Chow, 2002). It is composed of fructose (38%), glucose (31%), minerals, proteins, free amino acids, enzymes and vitamins (Pérez, 2002; Terrab et al., 2003). Also, a wide range of minor ingredients are present in honey, many of which are known to have antioxidant properties, such as phenolic acids (Dimitrova, Gevrenova, & Anklam, 2007; Martos et al., 2000; Tomas-Barberán, Martos, Ferreres, Radovic, & Anklam, 2001). Honey is rich in natural antioxidants and minerals. Thus, it takes a more active role to make a contribution to human health and nutrition. Honey has a wide range of colours from water white to dark amber or dark. Honey colour depends on various factors, mineral content being an important one. Light-coloured honeys usually have low ash content. On the contrary, dark-coloured honeys generally have higher ash contents (Al et al., 2009; Gomes, Dias, Moreira, Rodrigues, & Estevinho, 2010). The evidence of the biological actions of honey can be ascribed to its polyphenolic contents which, in turn, are ⇑ Corresponding author. Tel.: +98 2148292466; fax: +98 2148292200. E-mail address: [email protected] (S. Minaei). http://dx.doi.org/10.1016/j.foodchem.2014.02.136 0308-8146/Ó 2014 Elsevier Ltd. All rights reserved.
usually associated with its antioxidant and anti-inflammatory actions, as well as its cardiovascular, antiproliferative and antimicrobial benefits (Alvarez-suarez, Giampieri, & Battino, 2013). Some research groups have shown that the total phenol and flavonoids contents, and antioxidant activity of honey are greatly dependent on colour (Alvarez-Suarez et al., 2010; Bertoncelj, Doberšek, Jamnik, & Golob, 2007; Estevinho, Pereira, Moreira, Dias, & Pereira, 2008; Ferreira, Aires, Barreira, & Estevinho, 2009). Therefore, colour is an important feature in quality control of honey. The quality parameters (i.e., antioxidants) of honey are normally measured using conventional analytical techniques (Alvarez-Suarez et al., 2009). However, applying these techniques in honey industry has some disadvantages such as their destructive nature, implementation expense, and a time requirement. For food authentication and honey characterization in particular, introduction of alternative methods to conventional procedures aiming at objective measurement of honey in a consistent and cost effective manner is of great importance for the honey industry (Shafiee, Minaei, MoghaddamCharkari, Ghasemi-Varnamkhasti, & Barzegar, 2013). In this regard, several studies have been carried out to develop nondestructive techniques using a calibration data set obtained by analytical methods for various agricultural products (Brosnan & Sun, 2004; Butz, Hofmann, & Tauscher, 2005; Pace et al., 2013; Ruiz-Altisent et al., 2010). Among these approaches, computer vision can be noted for having good capability of food colour assessment as well as food characterization for those properties associated with
144
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
colour. Many research groups have used computer vision for analysis of food colour. The first step in the development of a computer vision technique for food quality assessment based on colour is colour calibration and characterization. Briones and Aguilera (2005) employed image analysis techniques to analyse colour variations during blooming development in chocolate. They concluded that computer vision systems have the capability to quantify overall changes as well as particular features over the whole chocolate surface. Thus, they introduced their vision based approach for customization and standardization of quality assessment of chocolate. Zheng et al. (2011) proposed a method to predict the changes of anthocyanins, ascorbic acid, total phenols, flavonoids, and antioxidant activity of red bayberry juice using CVS and ANN combination. In a recent study, CVS has been utilized to predict antioxidant activity and total phenols in pigmented carrots (Pace et al., 2013). It was demonstrated that CVS can act as an effective colour assessment tool and act as a proper predictor for antioxidant properties of carrot pigments. On the other hand, there is no report that employs CVS for predicting honey antioxidant activity, total phenolic content, and ash content. The objective of this research was to study the relationships between TPC, AA, AC, and honey colour. Moreover, we aim to evaluate the effectiveness of CVS–ANN combination for honey colour assessment and prediction of its TPC, AA, and AC. 2. Materials and methods 2.1. Honey samples and chemicals One hundred twenty nine honey samples of various floral origins and colours were analysed. These included: Loco (Astragalus bisulcatus), Opoponax-Tree, Alfalfa (Medicago sativa), Barberry (Berberis vulgaris), Thyme (Thymus vulgaris), Argentine thistle (Eryngium billardieri), and Dill (Anethum graveolens dhi). Samples were obtained directly from bee keepers of several different provinces of Iran. All chemicals and solvents used for analysis were analytical grade with the highest purity available. Folin–Ciocalteu’s phenol reagent, gallic acid, sodium carbonate anhydrous, sodium acetate trihydrate, ferrous ammonium sulphate, and ferric chloride hexahydrate were purchased from Merck (Merck KGaA, Darmstadt, Germany). Furthermore, 2,4,6-tripyridyl-s-triazine (TPTZ) was purchased from Sigma–Aldrich Chemie GmbH (Steinheim, Germany). 2.2. Colour measurement using Hunter lab and determination of ash content (AC) Before colour analysis, samples were heated at 50 °C for 10 min to dissolve sugar crystals. Then, they were paved and pureed into 5 cm Petri dishes at 1 cm height. Colour characteristics were measured using the CIE L⁄a⁄b⁄ colourimetric method. Honey colour values were determined using a commercial colourimeter (Hunter Lab, color Flex, USA). Percent ash content was determined for all honey samples based on AOAC (1990). 2.3. Total phenolic content (TPC) The Folin–Ciocalteu method was utilized to determine the total Phenolic content of the samples (Alvarez-Suarez et al., 2010). Honey solution in distilled water was prepared (10% w/v) and 0.5 ml of this solution was poured into a test tube. Then, 2.5 ml of Folin–Ciocalteu 0.2 N was added and well mixed. After 5 min, 2 ml of 0.7 M sodium carbonate was added. This solution was incubated in dark room at 25 °C for 2 h. Absorbance of the reaction mixture was measured at 760 nm against the sugar analogue using
spectrophotometer (Jenway, model 6505). Gallic acid was used as the standard to produce the calibration curve (50–250 mg/L). The total Phenolic content was expressed in mg of Gallic acid equivalents (mg GAE/kg of honey). 2.4. Quantification of antioxidant activity (AA) Several methods for determining the anti-oxidative activity in honey have been employed. The techniques to evaluate antioxidant capacity are based on colourimetric assays such as DPPH, FRAP, TEAC (ABTS) and microplate fluorescence reader like ORAC assay (Alvarez-Suarez et al., 2009). Antioxidants stabilize free radicals by donating electrons and the antioxidant capacity provided by this mechanism is determined by measuring the reduction capacity of metal ions such as ferric ion (Fe3+): FRAP (Niki, 2011). In this study, FRAP assay has been used for determination of honey antioxidant capacity (ferric reducing/antioxidant power). It is a simple direct test that widely used for assessment of antioxidant activity of various materials including honey (Bertoncelj et al., 2007). 2.4.1. The ferric reducing antioxidant power assay Benzie and Strain (1996) procedure with minor modification was used for FRAP assay. The principle of this method is based on the reduction of a ferric 2,4,6-tripyridyl-s-triazine complex (Fe3+-TPTZ) to its ferrous, coloured form (Fe2+-TPTZ) in the presence of antioxidants. The FRAP reagent contained 2.5 ml of a 10 mM TPTZ (2,4,6-tripyridyl-s-triazine) solution in 40 mM HCl, 2.5 ml of 20 mM FeCl3 and 25 ml of 0.3 M acetate buffer, pH 3.6. It was prepared daily and warmed to 37 °C. Aliquots of 400 ll of sample were mixed with 3.6 ml of FRAP reagent and the absorbance of the reaction mixture was measured at 593 nm after incubation at 37 °C for 10 min against the sugar analogue. Ammonium ferrous sulphate (100–1000 lM) was used for the calibration curves and the results were expressed as lmoles of ammonium ferrous sulphate per 100 g of honey (lmol Fe (II)/100 g of honey). All tests were carried out in triplicate and their averages were reported. PASW statistics software was utilized for statistical analysis. 2.5. Computer vision system (CVS) The computer vision system used in this work consists of four major components: dark chamber and lighting system, digital camera, computer hardware and software. A 30 30 45 cm3 wooden box was prepared to act as the housing. The bottom part of this box (15 cm) was used to form the illuminating system. Furthermore, the upper 30 cm was used to form the dark chamber. A plastic light diffuser sheet which is used to hold the sample separates these two sections from each other. Back-lighting was used for illumination of honey samples to diminish the negative effect of reflectance and to create homogenous light intensity on the honey sample surface. The back-lighting system consisted of a 30 30 15 cm3 box with its internal surfaces covered using a reflective sheet for more light scattering in the lighting chamber. LED lamps were placed horizontally in the chamber. A plastic light diffuser was placed on top of the chamber. A Canon 550 D kiss x4 camera was mounted vertically above the back lighting chamber at a distance of 20 cm. The camera was connected to the USB port of a Laptop (Dell, Inspiron 4030, China) provided with a remote capture software (Zoom Browser Ex. Version: 6.7) to view and to capture images from the camera setting controller. The internal surfaces of the wooden box were covered with dark sheets to prevent reflection and external light. The Honey sample was located against the center of the background (Fig. 1a). A standard grey card (10 12 cm2) with 18% reflectance (Kodak, USA) was used to set
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
145
Fig. 1. The scheme designed for prediction of the chemical attributes of honey using computer vision system and artificial neural networks.
the white balance of the CCD camera. Various different camera settings (combination of aperture, shutter speed and ISO) was tried to arrive at the average L⁄ pixel value and standard deviation of 48.1 ± 0.2 for the grey card. In order to achieve the illumination system stabilization, it was switched on 30 min prior to image acquisition. Images were captured using the following camera settings: exposure mode: manual, shutter speed: 0.3, lens aperture setting: 5.6, image resolution: 5184 3486 pixels, picture style: standard, and colour space: RGB. Camera was manually focused and the zoom was adjusted. Captured images were stored in bitmap format. All images were captured in triplicate.
pixels) was extracted from the center of each honey image (Fig. 1c). Bubbles and bee wax were removed from each image during the preprocessing step. Otsu thresholding method was used for separation of particles from the background (Fig. 1d). Then, the resulting image was subtracted from the original image. Finally, voids in the resulting image were filled with their neighbourhood intensities (Fig. 1e). This is plausible, because of the homogenous nature of the honey surface and close resemblance between the image intensities of neighbouring pixels. The RGB values for each colour patches were calculated by averaging RGB values of all pixels of processed cropped image.
2.6. Image processing
2.7. Colour transformation
The image preprocessing steps are presented in Fig. 1. After image acquisition and storage (Fig. 1b), crop image (1751 1275
Various colour spaces have been introduced within the last decade. Each of these colour spaces are formed by a linear or
146
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
non-linear transformation of RGB colour space. Because of their specific characterization, HSV, YIQ, CIE L⁄a⁄b⁄, and YCbCr transformations have been implemented in this study. 2.7.1. YIQ colour space A colour space known as YIQ is used in NTSC1. One of the main advantages of this format is that the intensity component is separated from colour data, such that the same signal can be used for both colour and black and white sets. In NTSC colour space, image data consists of three components: luminance (Y), hue (I), and saturation (Q). The first component represents greyscale information, while the last two components represent chrominance (colour information). To obtain YIQ values, the following transformation matrix is used:
2
3 2 3 2 3 Y 0:299 0:587 0:114 R 6 7 6 7 6 7 4 I 5 ¼ 4 0:596 0:274 0:322 5 4 G 5 Q 0:211 0:523 0:312 B 2.7.2. YCbCr colour space This colour space is widely used in digital video systems. Accordingly, luminance information is stored as a single component (Y), and chrominance information is stored as two colour-difference components (i.e., Cb and Cr). Cb represents the difference between the blue component and a reference value. Cr represents the difference between the red component and a reference value. The YCbCr transformation matrix is given below (Du & Sun, 2005):
2
3 2 3 2 3 Y 0:2989 0:5866 0:1145 R 6 7 6 7 6 7 4 Cb 5 ¼ 4 0:1688 0:3312 0:5000 5 4 G 5 Cr 0:5000 0:4184 0:0816 B 2.7.3. CIE L⁄ a⁄ b⁄ Based on CIE (1931) XYZ, CIE L⁄a⁄b⁄ was developed to measure the colour differences consistently with the perceived colour differences. The L⁄ channel stands for brightness, while a⁄ and b⁄ channels define the content of colour information, in which a⁄ extends from green to red and b⁄ from blue to yellow. It has been found that the Euclidian distance in L⁄a⁄b⁄ space provides better match to the human visual perception of colour. 2.7.4. HSV colour space HSV stands for hue, saturation, and value. It was introduced with the aim of providing a more intuitive colour mixing model. HSV attempts to capture the colour components, the way humans perceive colour. Furthermore, it can be found by using a simple transformation which makes it possible to use in real time applications. 2.8. Artificial neural networks (ANN) Colour calibration is concerned with adjusting the imaging system to a known state producing consistent and reproducible colour values. Characterization is the relationship between image primaries (RGB) and a device-independent colour space, such as the CIE XYZ or CIE LAB (Johnson, 1995; Westland & Ripamonti, 2004). In this sense, one way to use a digital camera as a colourimeter is to find a model, often some type of regression which directly maps RGB values to CIE XYZ or CIE LAB (Valous, Mendoza, Sun, & Allen, 2009). In this study, artificial neural networks technique was utilized for characterization of honey samples, and RGB mapping to CIE L⁄a⁄b⁄. Also, the prediction of honey TPC, AA, and AC were 1
National television systems committee.
done by ANN. A network with a single hidden layer was used to characterize honey colour. In different studies it has been shown that a single hidden layer would be sufficient for ANN to approximate any complex, nonlinear function (Dogan, Demirpence, & Cobaner, 2008). In this study, two feed-forward networks with input, output, and hidden layers were used for colour characterization and TPC, AA, and AC prediction. The Input layer of the network for honey colour measurement consisted of three neurons which corresponded to R, G, and B values extracted from honey sample images. The output layer had one neuron representing L⁄, a⁄, and b⁄ values of honey samples. Moreover, the network for TPC, AA, and AC predictions had 15 neurons in the input layer corresponding to 15 colour features extracted from each honey image (R, G, B, H, S, V, Y, I, Q, Y, Cb, Cr, L, a, and b) and one neuron in the output layer representing of TPC, AA, and AC of honey samples. Backpropagation algorithm was utilized in training of the ANN models. Because of the choice of the back-propagation network, the input data was normalized so that it ranged from zero to one in value before being fed to the ANN. Tangent-sigmoid and log-sigmoid transfer functions were used in the hidden layer and the output layer, respectively. Error minimization was accomplished using the Levenberg–Marquardt (LM) algorithm. A trial-and-error approach was utilized to set the number of neurons in the hidden layer. The Optimum number of nodes in the hidden layer was defined based on the lowest root mean square error value, RMSE. In this regard for any number of nodes in the hidden layer, execution of the ANN algorithm was repeated 10 times and the average RSME was calculated. For comparison of the actual values of CIEL⁄a⁄b⁄measured by the colourimeter and those predicted by the developed ANN model, colour differences were calculated. The average values of CIEL⁄a⁄b⁄ colour differences for validation and test sets was referred to as generalization error. The value of colour difference, DEab (CIE, 1986) was computed using the following equation:
DEab ¼
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 2 ðDL Þ þ ðDa Þ2 þ ðDb Þ
ð3Þ
The required codes were developed using MATLAB 2012 software. A schematic diagram of all stages for prediction of AA, TPC, and AC by the CVS and ANN combination is shown in Fig. 1. 3. Results and discussion 3.1. Honey colour and its relationships with TPC, AA, and AC Colour and chemical attributes of honey samples were determined as shown in Table 1. Colours of the honey samples were noticeably different and varied from white to dark amber. The brightest is Alfalfa honey that has the white colour to extra light amber tones. Moreover, the darkest one is Barberry honey that has the amber to dark amber colour tones. The mean value and standard deviation of L⁄, a⁄, and b⁄ colour parameters of honeys obtained using Hunter lab colourimeter are summarized in Table 1. Alfalfa, Loco, and Opponax samples had the highest average values of lightness indicated by L⁄ parameter (57.30, 56.23 and 54.22, respectively). There is no significant difference between these three types of honey and visual comparison confirms this. The dark amber Barberry honey was found to be the darkest having minimum value of lightness (20.06). Dill and Thyme honey are amber. In a report by González-Miret, Terrab, Hernanz, FernándezReca-males, and Heredia (2005) honeys were classified into two groups based on their lightness value (light honeys where L⁄ > 50) and dark honeys (where L⁄ < 50). In this regard, Alfalfa, Loco, Opponax and Argentine thistle are light honeys while Barberry, Dill and Thyme honeys are categorize as dark. Other colour parameter, a⁄ and b⁄ values, varied from 20.51 to 25.07
147
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150 Table 1 Mean ± STD values of colour parameters (L⁄a⁄b⁄ measured by Hunter Lab colourimeter), TPC, AA, and ash content of honey samples. Type of honey
Number of samples
L⁄
a⁄
b⁄
Ash content (%)
Total phenol content (mggallic kg)
1 2 3 4 5 6 7 Total
24 6 9 27 21 27 15 129
56.23 ± 3.23d 46.47 ± 3.56b 20.06 ± 1.03a 57.30 ± 3.46d 47.37 ± 4.87b 54.22 ± 3.92c,d 51.05 ± 5.98c
3.66 ± 1.38a,b 5.39 ± 1.32a,b 25.07 ± 1.11d 2.51 ± 0.89a 12.57 ± 3.01d 7.68 ± 1.76b,c 9.18 ± 1.57c,d
45.01 ± 4.06b,c 40.40 ± 7.78a,b,c 25.85 ± 1.31a 41.58 ± 6.30b 55.36 ± 4.68d 50.39 ± 7.93c,d 54.47 ± 3.58d
0.08 ± 0.02a,b 0.10 ± 0.03b 0.82 ± 0.01d 0.05 ± 0.02a 0.16 ± 0.02c 0.11 ± 0.06b 0.08 ± 0.01a,b
37.67 ± 5.53a 74.72 ± 7.63c 259.52 ± 34.49d 33.34 ± 5.96a 75.17 ± 5.11c 49.93 ± 10.2a,b 57.26 ± 11.45b,c
acid/
FRAP (lmol Fe (II)/ 100 g) 256.27 ± 16.11b 678.84 ± 41.87f 1383.18 ± 94.2g 204.14 ± 19.9a 558.02 ± 34.04e 325.28 ± 23.73c 431.37 ± 25.6d
Data with common letters have no significant differences (Duncan a < 0.05).
and 25.85 to 55.36 among the seven types of honey. There is a noteworthy difference between some types of honey for a⁄ and b⁄ parameters as shown in Table 1. Antioxidant activity of honey samples assessed by FRAP assay is entirely different (p < 0.05) among the seven types of honey. Barberry honey which is the darkest of all had the highest antioxidant activity and Alfalfa honey being the lightest honey had minimum antioxidant activity. Results showed that the TPC varied among different types of honey as its minimum value was for Alfalfa honey (33.34) which is the lightest. Moreover, the maximum value of TPC was associated with barberry, the darkest. The results confirmed the study done by Alvarez-Suarez et al. (2009) for Cuban honeys indicating that the darker samples includes higher AA and TPC values. Also, it has been found that the darker the honey, the higher the ash content. Thus the ash content of Barberry and Dill honeys are higher than that of the Alfalfa and Loco. This corroborates the results obtained by other researchers. Al et al. (2009) indicated that light-coloured honeys usually have low ash content while dark-coloured honey generally contain higher amount of ash. The relationships between honey colour and TPC, AA, and AC are presented in the correlation matrix below (presenting Pearson correlation coefficients).
2
L
a
Phenol content ash content
6 0:76 6a 6 6 Phenol content 0:78 0:73 6 6 0:70 0:64 0:56 4 ash content FRAP value
0:85 0:61 0:72
3 7 7 7 7 7 7 5
0:77
As it is found, all the relationships are statistically significant (p < 0.01). The highest correlation coefficients found is associated with the relationship between FRAP antioxidant activity and the L⁄ Value of honey colour (0.85), followed by TPC and L⁄ (0.78). These results confirm those obtained by Bertoncelj et al. (2007) for Slovenian honeys. Finally, The lower correlation (0.7) was found for ash content and L⁄ values of honey samples in comparison with FRAP value and total phenol content. The results of previous studies and this research show that honey colour is an important feature for honey characterization. Thus, it is possible to measure honey colour by a proper instrument and characterize honey for different colour-dependent parameters such as total phenol content, antioxidant activity, and ash content.
Table 2 Root mean square error of different number of nodes in the hidden layer for colour and chemical attributes of honey samples (mean ± standard deviation). Number of nodes in the hidden layer
1 2 3 4 5 6 7 8 9 10
RMSE L⁄
a⁄
b⁄
AC
TPC
AA
1.80 ± 0.48 1.83 ± 0.34 1.50 ± 0.11 1.59 ± 0.30 1.50 ± 0.16 1.73 ± 0.37 1.49 ± 0.54 1.91 ± 0.89 1.67 ± 0.34 1.42 ± 0.42
1.57 ± 0.40 1.60 ± 0.48 1.53 ± 0.34 1.65 ± 0.58 1.36 ± 0.29 1.48 ± 0.169 1.57 ± 0.630 1.41 ± 0.193 1.52 ± 0.665 1.38 ± 0.251
1.74 ± 0.14 2.15 ± 0.66 1.80 ± 0.30 1.99 ± 0.68 1.78 ± 0.66 1.90 ± 0.63 1.99 ± 0.93 1.47 ± 0.32 1.55 ± 0.28 1.38 ± 0.23
0.114 ± 0.019 0.109 ± 0.019 0.100 ± 0.015 0.107 ± 0. 01 0.107 ± 0.01 0.102 ± 0.01 0.091 ± 0.02 0.138 ± 0.03 0.080 ± 0.02 0.095 ± 0.01
4.02 ± 0.05 3.53 ± 0.38 3.09 ± 0.34 1.59 ± 0.21 1.56 ± 0.21 3.90 ± 0.41 3.31 ± 0.36 2.52 ± 0.31 1.65 ± 0.26 2.13 ± 0.22
7.19 ± 1.30 6.91 ± 1.63 5.46 ± 0.50 5.68 ± 0.72 5.88 ± 1.52 5.40 ± 0.97 6.62 ± 1.59 5.89 ± 1.77 5.44 ± 1.28 5.39 ± 0.82
Fig. 2. Correlations between the predicted and measured values of L⁄, a⁄ and b⁄ in the training set.
148
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
Fig. 3. Change in the values of DEab for each honey sample, considering the test and validation set of ANN.
3.2. Computer vision system and artificial neural network combination (CVS–ANN) The main objective of this study was to prove the applicability of CVS–ANN combination as a simple and potential tool for colour measurement and prediction of TPC, AA, and AC. In this regard, a proper computer vision system was developed and calibrated. Consistency of illumination system is an important factor, since colour reproducibility is dependent on this factor. A preliminary test was conducted for sensitivity analysis of the lighting system indicated the uniformity of illumination setup over the effective illuminated
area in the center of image. The value of DL⁄ for images captured in 20 min intervals during 2 h was lower than 0.09. This shows the good stability of illumination with time. The Homogeneously illuminated area of the grey card was a 8 12 cm2 rectangle in which the L⁄ value was at a consistent level of 48.1 (data not shown). After camera and lighting system calibrations, performing image processing, and colour transformation, ANN was used for colour characterization and TPC, AA, and AC prediction for honey samples. Network topology is an important factor in the design of ANNs since the topology type has a significant influence on the final network accuracy. The number of neurons in the hidden layer is one of the important factors in designing MLP networks with minimum error and maximum accuracy. In order to select the optimum number of neurons for the hidden layer, training step started with 1 neuron and was increased to 10 neurons. The original data set of 129 honey samples was divided into ten blocks, 70% of each block of data was randomly selected to generate the model. Data validation was performed with 15% of the total data (selected randomly from each block). In addition, the remaining 15% of data were used in the test phase. Table 2 shows the variation of RMSE in L⁄, a⁄, and b⁄ assessment, and the prediction of TPC, AA, and AC with various number of nodes in the hidden layer based on different colour features (15 different colour values for TPC, AA, and AC prediction). As shown in Table 2, the desired number of nodes in the hidden layer are 10, 5, 10, 9, 5, 10, respectively, for L⁄, a⁄, b⁄ assessments and, AC, TPC, and AA predictions.
Fig. 4. Correlations between the predicted and experimental values of AC, AA, and TPC (a, d, h) training set, (b, e, h) validation set, (c, f, i) test set.
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
Correlations between the experimental and predicted values of L⁄, a⁄ and b⁄ for the training set of honey samples are shown in Fig. 2. It can be seen that there is a proper correlations between ANN output and the actual measured values of honey colour parameters with R2 of 0.999, 0.999, and 0.974, and RMSE of 0.23, 0.14 and 1.55 for L⁄, a⁄ and b⁄. The CVS- ANN combination exhibited good prediction for the validation and test sets. Variation of, DEab for different honey samples (38 honey sample) in validation and test sets of ANN are shown in Fig. 3. The mean and standard deviation value for generalization error can be given as 1.01 ± 0.99. The minimum average distinguishable DEab by human eye has been reported to be 2.2 (Brainard, 2003). In this study, the generalization error average and standard deviation are below the threshold value of 2.2 (Fig. 3). This shows the effectiveness of the developed ANN model for honey colour measurement using CVS. Previous studies have shown the efficiency of CVS for food colour measurements. Valous et al., 2009 used colour calibration of a laboratory computer vision system for quality evaluation of pre-sliced hams. Their results showed that CVS can be used as an objective tool to specify colour measurement of ham slices. Thus it can be stated that, a CVS–ANN combination can be used for precise quantitative colour measurement of diverse food materials including honey. Correlations between the predicted and experimental AC, TPC and AA values for the training, validation and test sets are shown in Fig. 4 (a,d,g), (b,e,h) and (c,f,i), respectively. It can be seen that the optimal ANN could predict the AC, AA and TPC values of honey samples on the basis of colour data measured using CVS with reasonable R2 values of 0.99, 0.98, 0.87 and RMSE of 0.08, 1.58, and 5.4, respectively. These results show that colour parameters are good predictors for honey total phenol content, antioxidant activity and ash content. Furthermore CVS–ANN combination is a good tool for nondestructive measurement of honey nutrients. These results confirm those obtained by Zheng et al. (2011) for predicting anthocyanins and antioxidant activity of bayberry juice during storage. They reported that red, green, and blue intensity values are good predictors for anthocyanins and antioxidant activity parameters. Similar results were obtained by Pace et al. (2013) for pigmented carrots. They proposed two multiple regression models to predict AA and TPC of pigmented carrots on the basis of colour data measured using CVS. It was concluded that CVS provided more stable and informative colour evaluation, even for unevenly coloured surfaces compared to colourimeter. Our study results indicate that CVS–ANN can be effectively used for honey colour and nutrient assessment. 4. Conclusion This study was conducted for non-destructive characterization of honey using computer vision system. Since, ash content, antioxidant activity and total phenolic content of honey affect its colour; a computer vision system was used for prediction of these characteristics in honey samples using colour attributes of honey images. The proposed system measures honey colour with low generalization error. Thus, CVS can be used as a precise and fast colourimeter for honey colour assessment. Developed models for nutrition measurement of honey were able to predict AC, AA, and TPC with high regression coefficients. Our results show the accuracy and efficiency of CVS–ANN combination for non-destructive characterization of honey on a laboratory scale. A larger dataset — a large number of honey samples with various floral and geographical origins — can be used for programming the CVS which would be further use for industrial scale applications. This can be extended to other food products whose quality is associated with their colour, shape or texture.
149
Acknowledgemens The authors would like to express their appreciation to Dr. Mohammad Reza Khoshayand of Tehran University of Medical Sciences, Faculty of Pharmacy for providing the required substances for chemical tests. References Al, M. L., Daniel, D., Moise, A., Bobis, O., Laslo, L., & Bogdanov, S. (2009). Physicochemical and bioactive properties of different floral origin honeys from Romania. Food Chemistry, 112(4), 863–867. Alvarez-suarez, J. M., Giampieri, F., & Battino, M. (2013). Honey as a source of dietary antioxidants: Structures, bioavailability and evidence of protective effects against human chronic diseases. Current Medicinal Chemistry, 20(5), 621–638. Alvarez-Suarez, M., Tulipani, S., Díaz, D., Estevez, Y., Romandini, S., Giampieri, F., et al. (2010). Antioxidant and antimicrobial capacity of several monofloral Cuban honeys and their correlation with color, polyphenol content and other chemical compounds. Food and Chemical Toxicology, 48, 2490–2499. Alvarez-Suarez, J. M., Tulipani, S., Romandini, S., Vidal, A., & Battino, M. (2009). Methodological aspects about determination of compounds and in vitro evaluation of antioxidant capacity in the honey: A review. Current Analytical Chemistry, 5, 293–302. AOAC. (1990). Official methods of analysis of the association of analytical (15th ed.). Washington, DC: AOAC. Benzie, I. F., & Strain, J. (1996). The ferric reducing ability of plasma (FRAP) as a measure of ‘‘Antioxidant Power’’: The FRAP assay. Analytical Biochemistry, 239, 70–76. Bertoncelj, J., Doberšek, U., Jamnik, M., & Golob, J. (2007). Evaluation of the phenolic content, antioxidant activity, and color of Slovenian honey. Food Chemistry, 105, 822–828. Brainard, D. H. (2003). Color appearance and color difference specification. In S. K. Shevell (Ed.), The science of color (2nd ed., pp. 191–216). Amsterdam: Elsevier Science Ltd. Briones, V., & Aguilera, J. M. (2005). Image analysis of changes in surface color of chocolate. Food Research International, 38, 87–94. Brosnan, T., & Sun, D. W. (2004). Improving quality inspection of food products by computer vision – A review. Journal of Food Engineering, 61, 3–16. Butz, P., Hofmann, C., & Tauscher, B. (2005). Recent developments in non-invasive techniques for fresh fruit and vegetable internal quality analysis. Journal of Food Science, 70, 131–134. Chow, J. (2002). Probiotics and prebiotics: A brief overview. Journal of Renal Nutrition, 12, 76–86. CIE (1931). International Commission on Illumination. In Proceeding of the 8th session, Cambridge, England, UK. CIE (1986). Colorimetry – Official recommendations of the international commission on illumination. CIE Publication No. 15.2, CIE Central Bureau, Vienna, Austria. Dimitrova, B., Gevrenova, R., & Anklam, E. (2007). Analysis of acids in honeys of different floral origin by solid-phase extraction and high-performance liquid chromatography. Phytochemical Analysis, 18, 24–32. Dogan, A., Demirpence, H., & Cobaner, M. (2008). Prediction of groundwater levels from lake levels and climate data using ANN approach. Water SA, 34(2), 1–10. Du, C., & Sun, D. (2005). Comparison of three methods for classification of pizza topping using different colour spaces transformations. Journal of Food Engineering, 68, 277–287. Estevinho, L., Pereira, A. P., Moreira, L., Dias, L. G., & Pereira, E. (2008). Antioxidant and antimicrobial effects of compounds extracts of Northeast Portugal honey. Food and Chemical Toxicology, 46, 3774–3779. Ferreira, I. C. F. R., Aires, E., Barreira, J. C. M., & Estevinho, L. M. (2009). Antioxidant activity of Portuguese honey samples: Different contributions of the entire honey and extract. Food Chemistry, 114, 1438–1443. Gomes, S., Dias, L. G., Moreira, L. L., Rodrigues, P., & Estevinho, L. (2010). Physicochemical, microbiological and antimicrobial properties of commercial honeys from Portugal. Food and Chemical Toxicology, 48(2), 544–548. González-Miret, M. L., Terrab, A., Hernanz, D., Fernández-Reca-males, M. A., & Heredia, F. J. (2005). Multivariate correlation between color and mineral composition of honeys and by their botanical origin. Journal of Agricultural and Food Chemistry, 53, 2574–2580. Johnson, T. (1995). Methods for characterizing colour scanners and digital cameras. Displays, 16(4), 183–191. Martos, I., Ferreres, F., Yao, L., D’Arcy, B., Caffin, N., & Tomas-Barberan, F. A. (2000). Flavonoids in mono specific Eucalyptus honeys from Australia. Journal of Agricultural and Food Chemistry, 48, 4744–4748. Niki, E. (2011). Antioxidant capacity: Which capacity and how to assess it ? Journal of Berry Research, 1(4), 169–176. Pace, B., Cefola, M., Renna, F., Renna, M., Serio, F., & Attolico, G. (2013). Multiple regression models and computer vision systems to predict antioxidant activity and total phenols in pigmented carrots. Journal of Food Engineering, 117, 74–81. Pérez, R. A. (2002). Analysis of volatiles from Spanish honeys by solid-phase microextraction and gas chromatography–mass spectrometry. Journal of Agricultural and Food Chemistry, 50, 2633–2637.
150
S. Shafiee et al. / Food Chemistry 159 (2014) 143–150
Ruiz-Altisent, M., Ruiz-Garcia, L., Moreda, G. P., Lu, R., Hernandez-Sanchez, N., Correa, E. C., et al. (2010). Sensors for product characterization and quality of specialty crops – a review. Computers and Electronics in Agriculture, 74, 176–194. Shafiee, S., Minaei, S., Moghaddam-Charkari, N., Ghasemi-Varnamkhasti, M., & Barzegar, M. (2013). Potential application of machine vision to honey characterization. Trends in Food Science & Technology, 30, 174–177. Terrab, A., Gonzále, M. L., González, A. G., Díez, M. F., Escudero, M. L., & Heredia, F. J. (2003). Characterisation of Moroccan unifloral honeys using multivariate analysis. Europeans Food Research and Technology, 218, 88–95. Tomas-Barberán, F. A., Martos, I., Ferreres, F., Radovic, B. S., & Anklam, E. (2001). HPLC flavonoid profiles as markers for the botanical origin of European unifloral honeys. Journal of the Science of Food and Agriculture, 81, 485–496.
Valous, N. A., Mendoza, F., Sun, D., & Allen, P. (2009). Colour calibration of a laboratory computer vision system for quality evaluation of pre-sliced hams. Meat Science, 81, 132–141. Westland, S., & Ripamonti, C. (2004). Computational colour science using matlab. Chichester: John Wiley and Sons, pp. 111–140. Zheng, H., Jiang, L., Lou, H., Hu, Y., Kong, X., & Lu, H. (2011). Application of artificial neural network (ANN) and partial least-squares regression (PLSR) to predict the changes of anthocyanins, Ascorbic acid, total phenols, flavonoids, and antioxidant activity during storage of red bayberry juice based on fractal analysis and red, green, and blue (RGB) intensity values. Journal of Agricultural and Food Chemistry, 59, 592–600.
