An integrated defect detection method based on context encoder and perception-enhanced aggregation for cylinder bores

The detection of defects in cylinder bores is crucial for the industrial manufacturing of automobile engines. Current efforts which exhibit unstable accuracies, time inefficiencies and high-cost expenditures, have been mainly initiated by well-trained inspectors. In this paper, we build on the detection framework FasterRCNN to propose an integrated defect detection method based on context encoder and perception-enhanced aggregation for cylinder bores of automobile engines, named CBDetector. The improvements are twofold. The context encoder, which is a locally sensitive and globally covered integrated encoder composed of a CNN and an improved Transformer architecture used for well-rounded feature extraction, is proposed. To robustly perceive full scale cylinder bore defects, a perception-enhanced feature path aggregation unit is introduced. Extensive experiments conducted on our established dataset HIT-EngD and a public steel dataset NEU-DET demonstrate the SOTA performance of the CBDetector, with mAP50 increases of 22.7 and 7.8 compared to FasterRCNN integrated with Feature Pyramid Network on HIT-EngD and NEU-DET. Moreover, our method can run at a high frame rate (~10 FPS, Nvidia-A100).