Introduction: Pomegranate fruit as one of the most popular fruits native to Iran, belongs to Punica family (Punica granatum L). Iran with an annual production of about 700 tons is the largest producer of pomegranate fruits in the world. Colorfulness and healthiness are two important features of pomegranates, which cannot easily be controlled. Some negative characteristics of this fruit such as sun burning, cracking and scratchingcan reduce its economic value. Moreover, separating the arils from membrane (flesh) and sorting them based on their color and size is a laborious task which still is a challenging concern (Blasco et al., 2003). Despite these challenges, the demand for “ready-to-eat” of arils is increasing. Up to now several devices have been proposed to remove the arils from membrane with different operation principles. However, these devices leave some membrane segments with arils and also makeit difficult to sort the arils from color and size points of view (Khazaei et al., 2008; Singh et al., 2007). With the visual inspection methods, the external features of Bio-materials (e.g. shape, color and texture) can be evaluated. While for assessing their internal parameters, nondestructive methods such as MRI, X-RAY and NMR are preferred. To classify and identify bio-materials (e.g. fruits), several methods have been examined including Fuzzy technique (Hu et al., 1998), Multilayer (Luo et al., 1999) and Linear Discriminant Analysis (LAD) (Manickavasagan et al., (2010). The primary objective of this research wasto discriminate arils from membrane segments. Subsequently, the fruit components were classified into red, pink, white arils and membrane segments, using LAD method. Ultimately, the accuracy of classifications based on different images’ features was evaluated.
Materials and methods:Pomegranate fruits of Khazar variety were provided from Kashmar gardens. Prior to imaging step the fruits were categorized in four groups each of 50 samples. The arils were ranked as red, pink and white using human sensory. The images of arils samples were prepared using a Nikon Coolpix digital camera (Nikon co, Japan), in a chamber having six LED lamps, from a distance of 15 cm. During image processing, the images were first converted into grayscale format and then transformed into binary images. Subsequently, several morphological (see table 2) and textural image (see Table 3) features were extracted for classification purpose. For color features three color spaces including RGB, HSI and L*a*bwere examined (see Fig 3). The arils were classified and discriminated from membrane using 12 morphological, 10 color and six textural features. Linear Discriminant Analysis (LAD) was employed for classification based on the mentioned features. The validity of input data was examined using theleave-one-out cross validation method. Statistical analysis was carried out using SPSS ver. 16.
Results and discussion: The classification accuracy of arils based on morphological features was about 97.53% and the membrane segments were discriminated from arils with accuracy of 95.06% (Table 4). The classification with color features provided the accuracy of 45% when the “R” component of the images was considered (Table 5). This is mainly due to similar red band of the arils classes.The accuracy of classification improved whenHSI components were used andthe accuracy of 84% was achieved (Table 6). The best accuracy of classification with color features observed using L*a*b* color space. In this case the accuracy was 89.1% (Table 6). In the final stage of classification, six textural features obtained from statistical moments including mean grayscale, standard deviation, third moment, evenness, entropy and homogeneity were used. As shown in Table 7 with these components the accuracy of classification improved up 93.3%. Considering the classification with different features (morphological, color and textural) it can be said that, in general, the accuracy of discriminating membranes from arils is less accurate than the accuracy of discrimination between different arils (red, pink and white). This was observed in all methods of classifications with different image features. With regard to the specific functionality of each extracted feature, the combination of the features was used for classification. Due to the increasing number of input features, the stepwise method was used for rankingof input features.Out of 26 input features of classification model, 14 superior features were selected using stepwise method. The results of classificationwith the combination of different features are shown in Table 8. As it can be seen, the average accuracy of classification with the combination of features improved up to 99%. Fig. 4 shows the classification of the pomegranate components based on the combination of the features, using Linear Discriminant Analysis (LDA) method.
Conclusion: A classification model was employed to classify pomegranate arils and membranes, using Linear Discriminant Analysis method. To improve the accuracy of classification, different image features were extracted and examined. In order to achieve a higher accuracy, the combination of features wasalso tested. This improved the accuracy of classification up to 99%. Since the combination of features is a costly and time-consuming process, the stepwise method was used to rank and select the superior features before their use in classification step. |
