Lightweight YOLOv7 for bushing surface defects detection

Abstract

Bushings as metal structure parts because of their good welding performance, good plasticity strength, and other advantages in the engineering field are widely used. However, because the performance of the internal microstructure of the metal material is not uniform, there is an electrode potential difference between the micro area leading to surface corrosion; and the production process will inevitably produce defective surfaces with defective products, thus seriously affecting the subsequent use. Therefore, it is essential to accurately detect the defects on the surface of the bushing.At present, the inspection method based on machine vision has replaced the manual inspection method with low efficiency and a high false detection rate. However,due to the large amount of computation brought about by the complex network model, the efficiency can not meet the production needs of real-time detection; the simple network model is due to the limited ability to extract features and thus can not meet the requirements of the accuracy of the detection. To ensure the detection accuracy of the surface defects of the bushing and at the same time reduce the volume of the model, a bushing defect detection model based on the improved YOLOv7 is proposed. The backbone network of the model uses MobileNetv3 to replace the backbone network of the original YOLOv7, which improves the detection speed while guaranteeing the detection accuracy, and realizes the lightweight model at the same time; the CBAM (Convolutional Block Attention Module) attention mechanism is introduced into the residual edges of each layer of the backbone network, which pays more attention to the small-size target to get more important feature information; BiFPN feature pyramid is used to optimize the detection effect by weighted fusion of multi-scale feature information. The experimental results show that the improved algorithm in this paper reduces the mAP by only 0.7 percent compared with the traditional YOLOv7 algorithm; however, the detection speed is increased by 29.4 percent, and the model volume is reduced by 29.9 percent, which effectively improves the detection accuracy and speed of all kinds of defects on the surface of the bushings, and it can be better adapted to the industrial detecting environment. Mathematics Subject Classification (2020) 5206050