A Lightweight Network FLA-Detect for Steel Surface Defect Detection

Abstract Rapid and accurate defect detection is crucial for enhancing product quality in the steel industry. This paper proposes a steel surface defect detection network, FLA-Detect, to address these challenges. Three key improvements are integrated into FLA-Detect. To achieve a lightweight model, a new structure called FC2f has been introduced, significantly reducing the model's parameters Additionally, a lightweight attention mechanism named Large Separable Kernel Attention (LSKA) is incorporated, enhancing the accuracy of small target defect detection and improving network robustness. To further enhance feature extraction and model deployment feasibility, a new downsampling structure called AMDown is designed. The performance of FLA-Detect is validated through comparative ablation experiments on the NEU-DET dataset, followed by generalization validation on the GC10 dataset. These experimental results demonstrate the effectiveness and accuracy of FLA-Detect.