Car Brake Disc Surface Defect Detection Based on Improved YOLOv5

Abstract

Recycling and remanufacturing used automobile brake discs is an important measure to reduce environmental pollution and reduce production and manufacturing costs. Previous recycling detection methods mainly relied on manual visual inspection, which had problems such as low inspection efficiency and high missed detection rate. To address this problem, an improved YOLOv5 method for detecting surface defects on automobile brake discs is proposed. We choose the lightweight feature extraction network CloFormer to replace the original YOLOv5 basic network to improve the detection speed while reducing parameter complexity. At the same time, we introduce an improved GAN (Generative Adversarial Networks) network branch and use auxiliary classifiers to improve classification accuracy and reduce the missed detection rate. In addition, a genetic algorithm is used for bias optimization in the model fusion stage to improve the overall detection performance of the model. Afterwards, the improved model was compared and verified with similar algorithms on public data sets. Its recognition accuracy was significantly improved by 40.56%, and its mAP gain was 17.96%. It also achieved an FPS improvement of 38.72% and a parameter decline of 4.12%. Finally, when verified in real scenarios, the average accuracy rate was as high as 90.5%, the missed detection rate was 1.375%, and the false detection rate was reduced to 1.5%, achieving the expected results of the experiment. This method can provide a reference for the defect detection method of automobile brake disc recycling.