FAANet: Feature-Augmented Attention Network for Surface Defect Detection of Metal Workpieces

Accurate segmentation of defects on the surface of metal workpieces is challenging due to the large-scale differences and complex morphological features. To address the problem of low accuracy in segmenting metal workpiece surface defects in complex scenes, we propose a Feature-Augmented Attention Network (FAANet) by studying the types of workpiece surface defects. The network adopts an encoder-decoder structure. In the encoding stage, VGG16 is used as the backbone to extract features at different depths of the metal workpiece surface. Moreover, a feature augmentation module is designed to refine the feature maps at different levels for the characteristics of the information contained in the features at different depths, emphasizing the detailed features, such as the edges of the defective targets. In the decoding stage, an efficient channel attention mechanism is fused in the network to make full use of the feature map with rich information after feature fusion. More relevant features are extracted through the interaction of different channel information to improve information utilization as well as segmentation accuracy. For the sample imbalance problem, a loss function supervised training model incorporating focal loss is designed to mitigate the negative impact of sample imbalance on the model performance. The results show that the proposed method achieves better segmentation results in detecting surface defects on metal workpieces in complex scenes than existing prevailed methods.