摘要
针对现有白细胞图像语义分割模型中分割耗时长,以及过分割和欠分割导致分割精度低的问题,提出一种基于TM-DeeplabV3+的白细胞图像语义分割算法.该算法使用MobileNetV3网络替换Xception网络作为M-DeeplabV3+算法的主干特征提取网络,有效减少了语义分割模型的参数量;从源模型中选择预训练模型并进行参数迁移,进而通过迁移学习进行M-DeeplabV3+模型的迭代.在血液计数数据集中将TM-DeeplabV3+模型与DeeplabV3+模型进行对比实验验证,TM-DeeplabV3+模型的mIOU、mPA和分割速度指标分别提升了1.83%、1.45%和49.16帧/s.实验结果表明TM-DeeplabV3+算法明显提升了白细胞图像语义分割的准确性和实时性.
In view of the problem of long time of leukocyte images segmentation,low segmentation accuracy caused by over-segmentation and under-segmentation in the current existing semantic segmentation model,a semantic segmentation algorithm to leukocyte images based on Transfer MobileNetV3 DeeplabV3+(TM-DeeplabV3+)is proposed in the paper.In the algorithm,the MobileNetV3 network is been used to replace the Xception network as the backbone feature extraction network of MobileNetV3 DeeplabV3+(M-DeeplabV3+)algorithm to reduce the parameter amount of the semantic segmentation model.The pre-training model is been selected from the source model,the parameter transfer has been performed,and then the M-Deeplabv3+model has been iterated by transfer learning.The indicators of mIOU,mPA and running speed in the TM-DeeplabV3+model have been increased by 1.49%,1.45%,49.16 frames/s respectively based on the blood cell count dataset.The experiment results show that TM-DeeplabV3+algorithm has improved the real-time and accuracy of leukocyte semantic segmentation significantly.
作者
陈静
谢鹏
张亚飞
FELIX Manirankunda
CHEN Jing;XIE Peng;ZHANG Yafei;FELIX Manirankunda(School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan 232000,China)
出处
《湖北民族大学学报（自然科学版）》
CAS
2022年第3期314-319,335,共7页
Journal of Hubei Minzu University:Natural Science Edition
基金
国家自然科学基金项目(51874010)
安徽省教育厅自然科学基金项目(KJ2018A0087)
安徽理工大学研究生创新基金资助项目(2021CX2074).
