Defect detection in vehicle mirror nonplanar surfaces with multi-scale atrous single-shot detect mechanism

Abstract

Surface quality inspection is important for vehicle rearview mirrors. Surface quality defects, such as bubbles, particles, cracks, and scratches, may appear during the production process. Traditionally, manual inspection is time-consuming, laborious, low in accuracy, and costly. In recent years, object detectors based on deep learning have been used for defect detection [such as single shot detectors (SSDs)], and object detectors mostly detect the details of small objects at a shallow level. However, the lack of shallow semantic information will lead to inaccuracy. The deep layer has more semantic information, but the deep layer cannot be detected due to the effect of the complete convolutional layer in the convolutional neural network. In this article, we propose an enhanced SSD method to detect micro-defects on the nonplanar surface of vehicle rearview mirrors. We call it the multi-scale atrous single-shot detector (MSASSD). Specifically, we first replace the maximum pool depth layer with an unconscious convolutional layer to expand the receiving field without reducing the size of the input image. Then, we link the shallow layer to the deep layer through the fusion block to form new and rich fusion features for object detection. Finally, we use multi-scale features (including deep features and fusion features) to predict defects. The results show that our MSASSD method can improve the average accuracy of defect detection (about 1.2% compared with the SSD method), while the detection speed is equivalent (low about two frames per second compared to the SSD method).