A novel object detection algorithm based on Swin Transformer

This paper proposes an object detection algorithm based on Swin Transformer to address the issue of missed or false detections in the classic Faster R-CNN algorithm due to its difficulty in fully extracting global contextual features from images. Firstly, this paper adopts Swin Transformer to replace ResNet as the backbone feature extraction network in Faster R-CNN, aiming to enhance the network’s ability to perceive global information. Next, a multi-branch feature fusion module is designed. By introducing cross-scale connections between input and output feature maps, it enhances information transmission among feature maps of different scales, effectively alleviating dilution issues during information propagation along longer pathways. Finally, a Scale Perception Module (SPM), developed by improving the attention model, reduces information loss during nonlinear transformations, thereby strengthening the model’s feature utilization capability. Additionally, embedding the SPM into the multi-branch feature fusion module effectively enhances the model’s scale robustness and feature reuse ability. The proposed method achieved mAP and AP scores of 85.6 and 51.7 on the PASCAL VOC and MS COCO datasets, respectively. Furthermore, compared to other mainstream detection algorithms, it demonstrated improved detection accuracy across various categories and object scales, while maintaining excellent performance in recognizing small-scale objects.