Heterogeneous attention multi-scale network for efficient weld seam classification

Weld seam classification in industrial settings faces critical challenges including diverse weld geometries, subtle inter-class variations, and varying image quality under industrial conditions. This paper presents HAMS-Net (Heterogeneous Attention Multi-Scale Network), a novel deep learning framework that achieves state-of-the-art performance in weld seam classification. Our key innovation lies in the synergistic integration of four components: a Channel-Spatial Attention module for precise feature focus, a Heterogeneous Attention Pooling module for multi-scale feature capture, a computationally efficient Ghost Feature Channel ReLU layer, and an Adaptive Feature Pyramid Network for dynamic feature fusion. Extensive experiments demonstrate HAMS-Net's superior performance across ImageNet and specialized weld datasets compared to current benchmarks like Swin-Transformer and ViT. Notably, HAMS-Net maintains robust performance across varying industrial conditions while requiring fewer computational resources than existing methods. These results establish HAMS-Net as a practical solution for industrial weld classification, offering both improved accuracy and operational efficiency.