Abstract:
In the production process of aeroengine blades (AEBs), the surface defect detection of blades is substantial. Currently, most blade detection methods are aimed at large b...Show MoreMetadata
Abstract:
In the production process of aeroengine blades (AEBs), the surface defect detection of blades is substantial. Currently, most blade detection methods are aimed at large blades and are unsuitable for small blades. For the small blades, the defects are even tiny. In this article, we build blade optical detection equipment based on the flexible robotic arm to detect defects on the surface of AEBs during production. We also construct a dataset of surface defects on AEBs. Furthermore, we propose a cross-layer semantic guided network (CSGNet) based on YOLOv6 to detect tiny defects. In CSGNet, we introduce a cross-layer semantic guidance module (CSGM), which uses deep semantic information to guide the shallow feature layer to increase the detection performance of tiny defects. We design a furthest dynamic copy-paste (FDCP) data augmentation method by enriching the background information of samples and increasing the number of training samples dynamically to improve the detection performance of tiny defects. In addition, we optimize the detection head (AT-decoupled head) to improve the overall performance of the detector further. Experiments have proven that our network can achieve excellent detection results with less than 4{M} parameter amount, whereas AP can increase by 2.3% and {\text {AP}_{S}} can increase by 4.8% compared with YOLOv6. Generalizability experiments on the COCO dataset verify that our network still performs competitively on natural images.
Published in: IEEE Transactions on Instrumentation and Measurement ( Volume: 72)