Multi-scale Fusion Attention Network for Industrial Surface Defect Classification

With the development of industry, artificial recognition and traditional image processing are no longer sufficient to meet the current needs for surface defect detection of industrial products. Faced with various types of defects and low recognition efficiency, this paper proposes a multi-scale fusion attention network for effective and accurate identify defects such as leather. This network combines multiple kernels with different receptive field sizes and uses efficient group convolution to extract features of different scales, thus enhancing the ability to recognize different defects and improving recognition efficiency. In addition, the network dynamically selects multiple scale channel features based on attention mechanisms. We validate the effectiveness of our method on the NEU-CLS dataset and apply it to a collected leather defect dataset, conducting a series of comparative and ablation experiments. The experimental results demonstrate that our method not only has the highest accuracy but also is lightweight.