A Small-Sample Defect Classification Network and Interpretability Analysis

This paper tackles weld defect classification in X-ray films under small-sample constraints. We propose HRD-AlexNet, an enhanced network integrating a Hierarchical Residual Dense (HRD) module into AlexNet. Unlike standard dense connections, HRD employs cross-layer residual bridging and hierarchical aggregation, fusing early texture features with deeper semantics for multi-scale representation. Experiments show HRD-AlexNet outperforms baseline AlexNet, especially with limited data ($3.6 \%$ accuracy gain at $1 / 32$ training size). Using Focal Loss to address class imbalance, we achieve $88.66 \%$ accuracy. SHAP analysis enhances interpretability. The framework offers a robust, transparent solution for small-sample defect classification.