Grading zahedi dates based on external features using image processing and machine learning methods

Introduction: Date fruit is a strategic horticultural product in the Middle East that plays an important role as an economic product to develop exports. Iran is the second world producer that contains 14% production of date fruit in the world and has a high potential in order to ideally exploit this valuable product. Condidering the low price of Iran's exported dates due to poor preparation and packaging process it is necessary to use new technologies for classification and grading of them. The application of machine vision in agriculture has increased considerably in recent years.There are many fields in which computer vision is involved in order to develop precision agriculture. Machine vision systems by elimination of manual inspection in the field of postharvest technologies improve accurate and uniform quality control of agricultural products. In most of these applications, the method of image analysis for product categories, with the determination of some external features such as color, size, shape, and surface texture has been used (Blasco et al., 2012). Alohali used RGB images taken from Date fruits and defined a set of qualitative external features of dates and categorized them into three categories in terms of quality. One of the characteristics of the soft tissue was detected using color intensity distribution. The final precision of carefully designed system using a propagation neural network was 80% (Alohali, 2011). The size is a particular aspect of external appearance of fruits and vegetables; the price of agricultural products is usually related with their size; therefore, grading of fruits and vegetables into different size groups of size is always necessary in the postharvest handling and processing stages (Zhang et al., 2014). Texture is the other significant sensory quality attribute that has been frequently used in the external quality inspecting and grading systems for the agricultural product quality evaluation. Texture is closely related to some internal quality of fruits and vegetables, such as maturity and sugar content. Therefore, texture is one of the widely used indicators the consumer uses for quality assessment of fruits and vegetables. Texture analysis can also play an important role in defect recognition and segmentation in grading systems due to its powerful discriminating ability (Lee et al. 2008). Materials and Methods: The current study examined image processing technology for grading Zahedi cultivar dates in Khuzestan province. Each date fruit was placed under the camera and imaged. At the same time, the samples were classified by an experienced grader. Imaging was conducted in a lighting box to avoid the effects of ambient light. Capturing images was done by a digital camera using CCD sensor. External features of dates such as color, size, shape and surface texture were extracted by image processing methods using MATLAB software (Version R2013a, The Mathworks Inc., Natick, MA, USA). Eleven size and shape features, nine color features, and six external texture features were extracted. The features which led to better separablity for classification were selected using stepwise discriminant analysis (SDA). Selecting the best features is effective to increase accuracy and speed of the algorithm. Two methods of learning machine were used for final classification: discriminant analysis that is a statistical technique and neural networks (ANN). Discriminant analysis method and Neural networks were implemented in SPSS 22.0 and neurosolution 7.0 software respectively. Results and Discussion: The best channel to separate dates from background was identified by comparison of the histogram of 9 channels from RGB, HSI and Lab color spaces. The histogram graph, which has more breakdown in the range of intensities, is more suitable to apply thresholding operation because it has a good contrast with the background. Channel B from RGB color space was chosen to segment dates from background. Channel of B has a better contrast between the color channels and also its corresponding histogram intensity values led to the best separability. Five features of size and shape, three features of color and three features of external texture were selected by SDA method to reduce dimension of features space. Degrees marked from 1 to 3 for qualitative grading and sorting by size define levels of quality and determine size from big to small dates respectively. According to table 4, accuracy of classifications for grading, sorting by size and inspection of wrinkled date fruits from healthy ones were 93.6%, 94.4% and 90% respectively. Classifications by MLP neural networks were done. The most important factor for evaluation of neural networks is Correct classification rate (CCR%). The results based on CCR from Confusion matrix is reported in table 5. Accuracy of classifications for grading, sorting by size and inspection of wrinkled date fruits from healthy ones using ANN were 95.7%, 92.3% and 93.8% respectively. Conclusion: Final accuracy of classification using discriminant analysis and neural network was achieved 92.7% and 93.9% respectively. Results show relative superiority of neural networks over statistical methods due to its accuracy. According to high accuracy of classification using learning machine methods, it can be concluded that using image processing algorithm was successful in extracting external features for sorting and grading of dates.