Document Type : Short Article
Authors
1 Faculty of Agriculture, Ferdowsi University of Mashhad, Mashhad, Iran.
2 Department of Horticultural Science and Landscape Engineering, Hormozgan University, Iran.
3 Department of Horticultural Science and Landscape Engineering, Ferdowsi University of Mashhad, Mashhad, Iran
Abstract
Introduction: The diversity and abundance in quality properties of agricultural products are leading factors to develop non-destructive methods. Machine vision and artificial intelligence are powerful techniques in detection of many physical, mechanical and chemical properties of agricultural products. Prior to exporting, fruits are sorted in terms of their shapes, volumes or weights. Non-destructively taste-based sorting can be of importance in terms of markability and application. Artificial Neural Network (ANN) has been introduced as a new method to predict quality parameters such as firmness, total sugar content (TSC) and pH of agricultural products and to grade the products accordingly.
Material and Methods: In this research, the quality properties of Mango (Kelke- Sorkh var) were predicted using the combination of image processing and artificial intellect techniques. The mango samples were harvested from the orchard in Minab, Hormozgan province in Iran. The samples were transferred to computer vision lab, room temperature of 24°C and 22% RH. The samples were divided into three groups for temperature treatment. They were kept at three temperature levels of 5°C, 15°C and 24°C (control group) for 48 hours. The sample were then placed in room temperature and were imaged every second day for 14 day period. After imaging, each sample was undergon destructive tests for determining their quality attributes including sugar content, firmness and pH. The images were taken by a digital camera in visible spectrum (Nickon Coolpix p510, Nikon Inc, Japan). The taken images were, then, transferred to Matlab software environment (Mathworks Inc, US) for analysis and processing. The color factors from regions of intrest were extracted from the images in L*a*b* color space. The segmentation of images was performed by thresholding (threshhold value of 0.3) the image of difference between red and blue channels of taken RGB images. The conversion of RGB color space to L*a*b* was done by converting RGB image to XYZ basic color space first and before converting X, Y, and Z basic color components to L*, a*, b* color factors. The L* represent the lightness in the image from black (0) to white (100). In this project, a multilayer perceptron neural network with a hidden layer was used. The optimum number of neurons in the hidden layer was found to be 25. The maximum iterations was set as 1000 and the learning rate was set as 0.001.
Results and discussions: The input variables to the network were temperature treatment at three levels (control, 5°C and 15°C), the color factors (L*, a* and b*) and the variations of three color factors across the regions of interest (standard deviations of L*, a* and b*). The output variables were sugar content, pH and texture firmness. The results showed that the accuracy of the ANN model on the prediction of pH, sugar content and firmness were 45%, 85 and 88%, respectively. Although the accuracy of ANN model for predicting pH from color factors was rather low, this model was able to predict firmness and sugar content with highly accurately. The histogram of errors among three ANN models also showed the ANN model for predicting firmness and sugar content performed better than that for predicting pH. The MAPE prediction error were 9.53, 22.74 and 6.14, respectively, for predicting firmness, pH and sugar content. Comparing the results from the network in training and testing stages showed that ANN can be considered as a reliable method for estimating quality factors of firmness and sugar content with high accuracy and estimating pH with rather non-applicable accuracy.
Keywords
Send comment about this article