Blood Cell Count and Detection Method Based on YOLO

Abstract

Blood Cell Count and Detection (BCCD) has always been a popular topic in object detection and many researchers have applied and modified the two basic models: Faster RCNN and Yolo. However, it is still difficult to tell which model or modification would perform better on other BCCD datasets. Thus, this paper mainly focuses on finding a better model and modifications to BCCD example datasets containing 364 images of blood cells. Faster RCNN and Yolo v5 were used as the basic two models for the dataset. Through training and comparisons between the two models, the better model was chosen to make further modifications or adjustments to achieve a better maP result possible. The result shows that in this specific dataset, Yolo v5 performs better. The modified Yolo v5 model also has an improvement of 0.6 percent of map 0.5 and 0.5 percent of map 0.95 comparing to the original model, showing that modification of model configuration, model structures including head and backbone would efficiently improve the time taken for training and maP.