Potato Quality Grading Based on Depth Imaging and Convolutional Neural Network

As a cost-effective and nondestructive detection method, the machine vision technology has been widely applied in the detection of potato defects. Recently, the depth camera which supports range sensing has been used for potato surface defect detection, such as bumps and hollows. In this study, we developed a potato automatic grading system that uses a depth imaging system as a data collector and applies a machine learning system for potato quality grading. The depth imaging system collects 3D potato surface thickness distribution data and stores depth images for the training and validation of the machine learning system. The machine learning system, which is composed of a softmax regression model and a convolutional neural network model, can grade a potato tube into six different quality levels based on tube appearance and size. The experimental results indicate that the softmax regression model has a high accuracy in sample size detection, with a 94.4% success rate, but a low success rate in appearance classification (only 14.5% for the lowest group). The convolutional neural network model, however, achieved a high success rate not only in size classification, at 94.5%, but also in appearance classification, at 91.6%, and the overall quality grading accuracy was 86.6%. The quality grading based on the depth imaging technology shows its potential and advantages in nondestructive postharvesting research, especially for 3D surface shape-related fields.