A Self-Adaptive data balancing network for leukocyte detection based on CenterNet

Leukocyte detection in high-resolution large-scale blood smear images is a challenging task. Among them, leukocyte category imbalance is a real problem, in order to improve the accuracy of multi-category detection, this study starts from cell number imbalance, changes the traditional neural network inference, and constructs a full data-driven adaptive sampling method on Center Net benchmarks, and combines it with the DLA network architecture for the task of leukocyte multi-category detection on two high-resolution blood smear image datasets. Eventually, mAP achieves 94.1%, which is valuable for practical applications.