Pears classification by identifying internal defects based on X-ray images and neural networks

In order to increase the sales and profitability, it is essential to classify the pears according to the external morphology (including shape, color and luster) and internal defects that can be quantitatively detected by various approaches. However, the existing classification methods concentrate mainly on the external quality rather than the internal defects. Therefore, this investigation develops an efficient and accurate classification method that can identify the internal sclerosis and bruises by combining the X-ray non-destructive testing and the convolutional neural network. Initially, the relations between the characteristics of the internal defects, i.e., internal sclerosis and bruises, and the grayscale features of the X-ray images are analyzed to provide the experimental data and demonstrate the theoretical foundations. Then, the X-ray images are processed by resolution reduction, feature enhancement and gradient reconstruction to improve the training efficiency and classification precision. Finally, the 18-layer residual network (ResNet-18) is optimized and trained to identify the internal bruises and sclerosis and classify the pears based on the identification results. It is found that the overall accuracy can reach 96.67% for identifying the bruised and sclerotic pears. The proposed method could also be applied to other fruits for defects identification and quality classification.