Normalized Gaussian Wasserstein Distance and Vision Transformer Based Erythrocytes and Spirochaete Detection with Bi-Level Routing Attention and Soft-NMS

To solve the problem of detecting erythrocytes and spirochaete in blood, an improved YOLOV7 algorithm is proposed that integrates the attention mechanism, optimization loss function and NMS module, called SE-YOLOv7. This method used Normalized Gaussian Wasserstein Distance to optimize the original loss function, enhancing the model's detection performance for erythrocytes and spirochaete. Use soft-NMS instead of the NMS module to alleviate the network's detection problem of erythrocytes overlap. The addition of Bi-Level Routing Attention enhances the model's flexibility in computational allocation and content awareness. Experiments on DMID-ESD using Faster RCNN, SSD, Retinanet, YOLOV7 and SE-YOLOv7. Experimental results show that SE-YOLOv7 has a better detection effect, and its mAP is increased by 2.3% compared to YOLOv7. This is of great significance for achieving accurate detection of erythrocytes and spirochaete.