A study of measurement technology based on Structured Light Detection and Deep Learning

Abstract

This paper revealed an effective method to judge the quality levels of apples with optical images which could avoid the damages caused by touching measurement. When only using photographs of the apples, the quality levels cannot be effectively identified via auto optical inspection methods. We proposed a structured light projection system for detecting the apple surface topography which is corresponding to the small changes during storage periods. This system uses cosine waves with different phases to project onto the apple surface and then uses three phases to reconstruct the object. In order to get the most suitable spatial frequency for this system, we tested six spatial frequencies: f = 0.001, f = 0.003, f = 0.005, f = 0.008, f = 0.01, and f = 0.08. Test results show that the spatial frequency f = 0.008 is best for our samples and hardware equipment. A convolutional neural network modified from LeNet is used to classify both the pictures: directly captured from apples and the reconstructed pictures from the structured lighting system. Experiments were performed on 30 fresh apples which then storage for a period of 7 days, 14 days and 21 days. The results of the convolutional neural network trained on structured light system samples showed that the accuracy was 87%, but the accuracy of the results of only using photography of apples was just 46%.