Image Recognition With Image Enhancement and Swin Transformer Improved YOLOXs Algorithm

Aiming at the problem of insufficient performance of existing object detection models under small-object recognition, complex background interference, and lighting changes, this study proposes an improved YOLOXs model that integrates multi-stage image enhancement, coordinated attention mechanism, and hierarchical vision Transformer. This study enhances the diversity of training data through a multi-stage data enhancement strategy, introduces coordinate attention mechanism to strengthen the key feature expression. Moreover, the global modeling capability of Swin Transformer is utilized to enhance the long-distance target association. According to experimental results, the enhanced model outperformed the original YOLOXs by 5.6%, achieving a mAP@0.5 of 46.8% on the COCO 2017 test set. In the small-object detection task, mAP@0.5 for small-objects improved to 36.7%, an increase of 8.2%. On the self-built Small-Object Dataset, the detection accuracy of the improved model reached 39.4%, an increase of 9.3% compared to the baseline model. In addition, in real-world scenarios such as traffic surveillance, unreturned aerial vehicle aerial photography, medical imaging, and industrial inspection, the average IoU of the improved model increased by up to 7.2%. The miss rate and false alarm rate were reduced by 4.4% and 2.5%, respectively. Overall, the proposed method can significantly enhance the robustness and accuracy of the object detection model in complex scenes. It also provides an effective solution for recognizing small-objects and performing vision tasks in low-contrast environments.