A Metal Surface Defect Detection Algorithm Based on Mixed Supervised and Cross Stage Partial Darknet

In response to the limitations of traditional visual identification detection methods, such as low accuracy of metal structure surface damage detection and high dependence on labeled data, a mixed supervised model based on cross stage partial (CSP) Darknet is proposed. By utilizing the CSPDarkNet for feature extraction, the capability of model feature extraction is enhanced while meeting lightweight requirements. Additionally, the architecture introduces a feature pyramid structure to fuse different scale feature layers in order to strengthen network representation capability. Furthermore, a non-local attention mechanism is employed to enhance the contextual modeling ability in image processing. Meanwhile, the use of a mixed supervised training approach is explored, and the impact of pixel-level label proportions on detection accuracy is investigated. As a result, high accuracy is achieved in the industrial inspection datasets KolektorSDD, meeting practical requirements.