A Method for Improving Urinary Sediment Object Detection of YOLOv5

In order to accurately and quickly detect urinary sediments under diverse background, in this paper, we propose a model for improving urinary sediment detection of YOLOv5 detector and name it ATT_YOLOv5, which mainly combines with CBAM attention in backbone, and it thus could solve the problem of losing semantic and global features caused by convolutional option. Meanwhile, inspired by the way about attention's weights extraction and the structure of adaptively spatial feature fusion, we propose a new detection head module about adaptively all feature share, which names AAFS and it efficiently restrains background information. In order to solve the problem of the lack of medical urine sediments dataset, we construct a standard dataset of urinary sediment, which includes 12 categories and more than 30000 images. We get the best detection result 78.2 % at AP@.5 on the dataset. It is almost 5% higher than YOLOv5 detectors, and the detection performance on objects in almost scale is improved.