A Steel Surface Defect Detection Algorithm Based on Improved YOLOv7

Abstract

To address the issue of low detection accuracy due to poor-quality steel surfaces, a complex background, and varying defect sizes, this study presents a novel approach to the identification of flaws in the surface on steel by implementing the YOLOv7 framework. Initially, with the BoTNet feature added to the basic structure, the method's base gets better at what it does and can gather information. Subsequently, the nearest neighbor interpolation used in the sampling mechanism of the head network is exchanged for the CARAFE slim upsampling tool to improve its functionality integration qualities. Lastly, the prediction head section adopts the enhanced shuffled attention mechanism (ESAM) to make the algorithm better at making predictions. In the final phase, the algorithm attains a mAP of 84.7% on the NEU-DET dataset and a detection speed of 65 FPS.