AF Det: An Attention Fusion Convolutional Network for High Fidelity Industrial Defect Detection

In modern industrial manufacturing, automated surface defect detection is a key link to ensure product quality and improve production efficiency. However, traditional detection methods based on convolutional neural networks often face two major challenges in processing industrial images: firstly, defect features are usually extremely subtle and have huge intra class morphological differences, which require extremely high feature recognition ability of the model; Secondly, in the context of a massive number of qualified products, the proportion of defective samples is extremely low, which makes it difficult for the standard training objective function to effectively learn sparse defect features. To address these challenges, this paper proposes an end-to-end defect detection framework called Attention Fused Detector (AF Det). The core of this framework lies in the dual optimization of “attention” resources during network training and inference processes. Firstly, by seamlessly embedding convolutional block attention modules in the deep feature extraction backbone network, AF Det can adaptively recalibrate the spatial and channel dimensions of the feature map, thereby enhancing its ability to characterize key defect areas. Secondly, in order to address the learning bias caused by class imbalance, we designed and adopted a mixed loss function that combines Focal Loss designed specifically for handling difficult to distinguish samples with Smooth L1 Loss for robust bounding box regression, guiding the model to focus on “what to learn” from the optimization objective level. The simulation experiments conducted on the publicly available NEU-DET industrial surface defect dataset show that the proposed AF Det framework achieves an average accuracy of 81.5%, verifying the effectiveness of this method in improving the accuracy and robustness of industrial defect detection.