SN-TOOD: An Improved TOOD-Based Steel Surface Defect Detection

The quality of steel products is a critical determinant in the competitive landscape and growth of the steel industry. Despite this, the surfaces of steel products are inevitably marred by various imperfections due to a multitude of factors encountered in the manufacturing process. Consequently, the detection and analysis of these imperfections are paramount for improving product quality and enhancing the competitive edge of steel enterprises. In this context, the present study introduces an enhanced defect detection model based on Task-aligned One-stage Object Detection (TOOD), which exhibits superior performance in the identification of surface defects on steel plates, designated as SN-TOOD. Firstly, an Spatial and Channel Multi-scale Attention module is proposed to refine the model's spatial and channel-based feature representation of steel surface defects. Following this, feature pyramid networks augmented by neural architecture search is integrated as a feature fusion network, which autonomously identifies the optimal configuration for feature fusion pertinent to steel surface defect detection during the training phase. Furthermore, the soft non-maximum suppression is utilized in the post-processing stage to remove redundant bounding boxes generated during the model's predictive phase, thereby augmenting the detection accuracy. Finally, a comprehensive experimental validation has substantiated that the proposed methodology surpasses alternative approaches in terms of the mAP50 metric (NEU-DET: 0.712).