Accurate Location Detection Method for Aluminum Profile Surface Defects Based on Improved YOLOX-S Algorithm

Aluminum profiles in the production process will inevitably appear a variety of surface defects, seriously affect the quality of products. The traditional method to detect the surface defects can not meet the actual demand, so it is of great significance to study the efficient detection method. In this paper, digital image processing methods such as rotation, flip, contrast and brightness transformation are used to increase the number of samples and simulate the complex imaging environment. An improved YOLOX-S detection model is proposed. Squeeze-and-Excitation Networks is embedded in the Cross Stage Partial module, and then SECSP module is proposed, and all CSP modules in YOLO-S are replaced with SECSP module, which improves the sensitivity of the network to the feature channel. SCYLLA-IoU loss function is used instead of IoU loss function. The improved model can improve the detection ability of small targets and the ability to resist background interference information. The mAP reaches 91.62, which is 1.82% higher than that of the basic YOLOX-S, and the detection speed reaches 58.67 frames ·s⁻¹, which can meet the real-time detection requirements. At the same time, the comparison experiment proves that the comprehensive performance of the proposed detection model is the best, and the detection accuracy and speed have reached a good balance. The ablation experiment proves that the two improved schemes can improve the detection accuracy of the network.

Graphical Abstract