TE-YOLOF: Tiny and efficient YOLOF for blood cell detection. (March 2022)

• Record Type: Journal Article
• Authors: Xu, Fanxin; Li, Xiangkui; Yang, Hang; Wang, Yali; Xiang, Wei
• Is Part Of: Biomedical signal processing and control. Volume 73(2022)
• Publication Date: 2022-03
• Subjects: Blood cell detection -- YOLOF -- EfficientNet -- Depthwise separable convolution -- Mish; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103416
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We propose a new light-weight model based on YOLOF to solve the relatively low precision of red blood cell detection problem that the FED model faced. We make further light-weight improvements to YOLOF, reducing the model complexity to less than 10M and improving the performance of blood cell detection. For each component we used, we have done ablation experiments to prove its advantages. The proposed model TE-YOLOF can be generalized to other datasets for detection directly. It shows the great potential to achieve robustness in the field of blood cell detection. Abstract: Blood cell detection in microscopic images is an essential branch of medical image processing research. The research and application of computer vision algorithms in this field are more concerned about the trade-off between accuracy and model complexity. The FED detector modified by YOLOv3 is a representative light-weight model to detect blood cell objects such as red blood cells, white blood cells and platelets. But the detection precision of red blood cells in the FED model is relatively low compared with platelets and white blood cells due to the imbalance distribution of different types of cells. In this research, we propose a light-weight model based on YOLOF to address the relatively low precision of red blood cell detection problem, in order to achieve the overall improvement of detection precision. This object detector is called TE-YOLOF, Tiny and Efficient YOLOF. Model light-weighting is accomplished with the excellent feature extraction capabilities of EfficientNet as backbone and the ability of the Depthwise Separable Convolution to reduce the number of parameters while maintaining precision. Furthermore, the Mish activation function is employed to increase the precision. Extensive experiments on the BCCD dataset prove the effectiveness of the proposed model, which can achieve higher precision with less parameters than FED. TE-YOLOF is also effective on other cross-domain blood cell detection experiments.