Detection of Small Size Defects in Belt Layer of Radial Tire Based on Improved Faster R-CNN

The cord structure of belt layer in radial tire is complex, and various defects such as cord overlapping, cord cracking and impurities may occur during manufacturing. With the development of deep learning, object detection based on convolutional neural network become a common defect detection method. Due to the variety of defect sizes, the receptive field of the feature map output by this kind of method is large, whereas the small-size defect yields weak feature and may be readily missed. In order to solve the problem that the feature extracted by the convolutional neural network of small-size defect is weak, a belt layer defect detection method based on improved Faster R-CNN is proposed. Faster-R-CNN’s convolution layers are used for feature fusion to solve the problem of insufficient feature extraction for subtle weeny defects. At the same time, Distance Intersection over Union (DIoU) is used to obtain object box that’s more sensitive to object scale to solve the problem of loose defect bounding boxes. The algorithm steps are as follows: Firstly, the belt layer area is segmented. We first segment the shoulder and belt layer areas by vertical projection, and then combine extreme value filtering (EVF) with binarization to segment the belt layer area according to the characteristics of the horizontal cord in the shoulder area. Secondly, construct the defect dataset of the belt layer and enlarge the area proportion of the defect target in the image. Thirdly, the shared convolutional layer of Faster R-CNN is used for front-layer feature fusion to ensure the feature map include higher-level features and higher resolution features. Finally, DIoU is used to get a bounding box that is more scale-sensitive. Experiments were conducted on the defect dataset containing 6316 object boxes for training and 1036 object boxes for test. Compared with the vanilla Faster R-CNN, the false negative rate decreased by 7.79%, the false positive rate decreased by 3.4%, the f-score improved by 5% and the detection box is more fitting for the defect object.