AYOLOv5: Improved YOLOv5 based on attention mechanism for blood cell detection

Abstract

A crucial component of biological study and clinical trials is the microscopic examination of cells and tissues. Although tremendous progress has been made in computer-assisted microscopy cell detection techniques, especially recently based on deep learning, it still needs to be determined to accurately identify cell targets in the presence of dense and complex cell distribution. In this research, an improved YOLOv5 (AYOLOv5) based on the attention mechanism is suggested to address the issue of the low recognition rate of cell detection caused by this circumstance. Based on YOLOv5, the attention mechanism is introduced to improve the convolutional neural network's features in areas of the picture with a high density of features. The convolutional block attention module (CBAM) and the transformer encoder block are used in this study to develop the attention mechanism. The YOLOv5 integrated convolutional block attention module increases the weight of cell-dense regions in blood cell pictures and aids the network's ability to resist information other than cells. Additionally, the transformer block is introduced to YOLOv5′s processing of upper and lower feature data to improve the network's capacity to gather details about various cell properties, enabling AYOLOv5 to recognize and distinguish blood cells in cell-dense areas. The experiment was done on the dataset BCCD, and the mAP results for cell detection reached 93.3%, better than previously discovered. Also, the validation set's average recognition accuracy increased from 89% to 98%. The experimental results demonstrated that the suggested AYOLOv5 could extract the cells' feature information more effectively, considerably improving the cell pictures and recognition performance.