Performance Evaluation of Automatic Labeling System of Crack Length at Tooth Root and Examination of Erroneous Detection

Abstract

In this paper, a semi-automatic labeling system of crack lengths at the tooth root of plastic gears is developed. Captured side-view images of plastic gears by a high-speed camera have been manually labeled with the quantity of crack lengths. Based on the large amount of labelled image data obtained from previous research, the developed system using a deep neural network classifies the side-view images superadded by endurance tests of plastic gears according to the crack length, automatically. The system employs a pre-learned popular convolutional neural network called VGG16, which has 1000 classification ability. The system was modified to fit a four classifications problem related to the crack length (“no crack”, “~ 40%”, “~ 70%” or “~ 100%”), and the weight of two layers, which were close to the output layer, were relearned with transfer learning. As a result of the learning, the classification accuracy of the unknown images was about 90%, Finally, misclassified images were investigated, and the hard-to-classify images were revealed.