Multi-Scale Fusion and Dual Attention Enhanced EfficientNet with Class-Aware Multi-Branch for Steel Defect Classification-Related Results

Steel surface defect inspection is a critical component of industrial quality control, where traditional defect inspection methods suffer from limitations including low accuracy, poor efficiency, and dependence on manual inspection. This paper proposes a multi-scale feature fusion deep learning model based on EfficientNet-B4, termed MSFEffB4, for multi-label classification of steel surface defects. The model integrates feature information from different network hierarchies through a multi-scale feature fusion module and enhances representation capability of defect features using a dual attention mechanism. To address the prevalent problem of class imbalance in industrial datasets, an adaptive minority class augmentation strategy is designed, which incorporates intelligent data balancing and class-balanced loss functions to improve classification performance for minority defect categories. The experimental approach employs a two-stage training strategy, where the first stage performs feature learning using augmented data, and the second stage conducts fine-tuning optimization on original data. Experimental results on the Severstal Steel Defect Detection dataset demonstrate that the proposed method achieves a macro-averaged F1-score of 95.18%, a mean label accuracy of 98.57%, a mean AUC of 99.34%, and a mean average precision (mAP) of 98.17%, respectively. Five-fold cross-validation experiments further validate the stability and generalization capability of the method, thus providing a viable technical solution for industrial defect classification tasks.