Abstract:

Blood cells detection and counting are important task of blood examination. However, a large number of  cell microscopic image has various shapes, small object and complicated background, it is still a very challenge task to recognize the blood cell automatically. In order to detect the small object from complicated background more effective, a novel YOLO-Dense network model is proposed based on one-stage deep detection framework YOLO. Firstly, the proposed model uses [K]-means to cluster the size of anchor box, three different sizes of anchor box for potential detection object are obtained, then the residual network module and multiple-scale module based on feature pyramids are introduced to improve the recognition accuracy of small object. Further, the gradient dispersion and explosion for deep network training can be solved effectively through the skip-connection of the residual module. Finally, by integrating the dense-network module in the proposed model, the network inference speed is further improved. The experimental resuls show that the mAP and detection time of YOLO-Dense are 0.86 and 24.9 ms respectively. Comparing with Faster R-CNN and original YOLO algorithm, the proposed model has achieved best performance for blood cell detection.

Key words: blood examination, cell count, deep learning, object detection, YOLO

摘要：

血细胞检测和计数是血液检验中的一项重要内容，大量细胞显微图像形态多样、目标较小且背景复杂，自动识别血细胞仍然是一项具有挑战性的任务。为解决血细胞检测中复杂小目标识别问题，基于一阶段深度检测YOLO框架，提出了一种新颖的YOLO-Dense网络模型。通过使用[K]-means算法对锚框进行聚类，获得三种不同大小的潜在待识别目标的锚框，并在YOLO基础网络中引入残差模块和特征金字塔的多尺度模块，以提高对小目标的识别精度；通过跳层连接进行残差训练，有效解决深度网络梯度弥散和爆炸等问题；通过在网络架构中增加密集连接模块，使得提出的模型能够有效提升网络推断速度。实验结果表明：YOLO-Dense网络均值平均精度mAP和检测时间分别为0.86和24.9 ms。相比Faster R-CNN和原始的YOLO网络，YOLO-Dense模型在血细胞检测上取得了最好的性能。

关键词: 血液检验, 血细胞计数, 深度学习, 目标检测, YOLO