High Performance Target Detection Based on Swin Transformer

In modern computer vision tasks, Swin Transformer is an emerging deep learning model architecture that innovates on the basis of the traditional transformer architecture by the introduction of a mechanism of attention called Shifted Windows to improve the performance and efficiency of the model. This paper presents a novel target detection method based on the original YOLOv5 detection model. The method combines the multi-layer attention mechanism of Swin Transformer with adaptive features and introduces Swin Transformer as a backbone network to replace the original backbone network. On the one hand, swin transformer is used as the backbone network of the target detection, which adopts the multi-layer attention mechanism to perform self-attention operation on features with different scales, and divides the feature map of the target image through the window shift mechanism to achieve the fusion of multiple disjoint regions. On the other hand, swin transformer's adaptive ability in multi-scale and multi-level feature representation enables it to achieve more accurate data analysis, thus achieving high performance target detection and multi-scale perception.