SST-YOLOv5s: advancing real-time blood cell object detection through multi-headed attention mechanism

作     者：Zhang, Mingyu Chen, Jiaqing 

作者机构：Xiamen Univ Wang Yanan Inst Studies Econ Xiamen 361005 Peoples R China Wuhan Univ Technol Coll Math & Stat Wuhan 430070 Peoples R China 

出 版 物：《SIGNAL IMAGE AND VIDEO PROCESSING》 (Signal Image Video Process.)

年 卷 期：2025年第19卷第3期

页      面：1-17页

核心收录：

学科分类：0808[工学-电气工程] 1002[医学-临床医学] 0809[工学-电子科学与技术（可授工学、理学学位）] 08[工学] 

基　　金：National Natural Science Foundation Open Fund of Hubei Longzhong Laboratory 

主　　题：Deep learning Medical image process Blood cell YOLOv5s Object detection 

摘      要：An important method of diagnosing diseases in medicine is the examination of blood samples in which the blood cells in the blood sample are examined. To address the challenges of detecting small and variable blood cells in real-time, we introduce a small-target Swin Transformer YOLOv5s model named SST-YOLOv5s in this study. Based on YOLOv5s, the small object detection layer is added to enhance the feature extraction capability of YOLOv5s and improve the detection of small targets in the dataset. In this way, the C3 module is replaced by the Swin Transformer Blocks, which enhances the extraction of global information from the model as well as speeds up the computation of the model. Subsequently, considering the need for the testing speed of the model, Simplify Optimal Transport Assignment is used in this research to reduce the testing time of the model greatly. The experimental results demonstrate that compared with the baseline model of YOLOv5s, the SST-YOLOv5s model achieves a mAP0.5 of 93.0%, a mAP0.5:0.95 of 60.4%, and a Latency of 0.6. To the best of our knowledge, SST-YOLOv5s is the first YOLOv5s to add Swin Transformer and SimOTA to a small-target detection head and use it to detect cells and further proves that SST-YOLOv5s can be better applied in practical medical applications.