Multi-scale Feature Fusion Extraction Structure for the Leather Defect Detection Algorithm

Detection of surface defects using computer vision is a pivotal technology for achieving intelligent manufacturing. Leather products are one of the most widely traded goods in the world, and automatic identification, localization, and detection of surface defects in leather are indispensable for achieving intelligent manufacturing of leather products. This research paper proposes FG-DETR, a multi-scale feature fusion extraction structure for small target detection algorithms. Given the recent success of deep learning methods in various related fields and that certain defects and flaws in leather pictures in our dataset are relatively small. FG-DETR is based on DETR with some improvements as follows, Initially, DETR's ResNet50 network performs feature extraction on the original image and produces four scales of feature maps. Second, we apply a modified Feature Pyramid Network (FPN) to downsample and upsample the four scales of feature maps, followed by the fusion and output of a single feature map. A Gaussian kernel and output then process the fused feature maps as the final feature maps. Finally, the feature maps are merged with the position encoding information and combined into the Transformer encoder-decoder architecture, resulting in the detection outcomes through the Position-wise Feed-Forward Networks (FFN). In our leather dataset, the FG-DETR model is more effective in detecting small targets than DETR, as our leather images are large, and the original DETR is nearly ineffective for images with relatively small defects. The average detection accuracy of medium and large targets also improved, with a 7.6 percentage point improvement at AP50. The experimental findings demonstrate that the multi-scale feature fusion extraction structure significantly enhances the detection accuracy of DETR.