医学图像特征相似度度量的网络压缩研究  

Research on Network Compression for Similarity Measurement of Medical Image Features

作　　者：杨东海[1] 许淑茹[1] YANG Donghai;XU Shuru(Zhangzhou Health Vocational College,Zhangzhou 363000,China)

机构地区：[1]漳州卫生职业学院,福建漳州363000

出　　处：《通化师范学院学报》2024年第4期65-72,共8页Journal of Tonghua Normal University

基　　金：福建省中青年教师教育科研项目(科技类)(JAT210867,JZ180828);漳州市自然科学基金(ZZ2021J43);漳州卫生职业学院院级科研项目(ZWYZ202104);漳州卫生职业学院院级教学改革研究课题(ZWYJ202218)。

摘　　要：为提高深度学习网络迁移应用于医学图像领域的性能,针对通道间相互独立剪枝压缩网络导致模型性能下降的问题,采用峰值信噪比方法对网络层内部及网络层之间的特征激活图进行相似度度量,整体评价层内及层间通道的重要性.首先,对医学图像进行预处理,采用峰值信噪比法度量预处理后图像与网络的各节点汇合层的相似度确定迁移网络深度.其次,对同一层内特征和不同层间的特征进行特征相似度度量并按照度量结果进行排序.最后,通过特征度量结果和期望压缩系数对网络进行压缩.实验结果表明:采用特征相似度度量对迁移网络通道剪枝仅小幅降低性能就能有效进行网络压缩.因此采用峰值信噪比法度量特征相似度可以作为通道剪枝的判断依据.In order to improve the performance of application of deep learning and network transfer to the medical image field,the peak signal-to-noise ratio method was used to measure the similarity of feature activation graphs of layers,so as to evaluate the importance of intra-layer and inter-layer channels.Firstly,the medical image was preprocessed,and the peak signal-to-noise ratio method was used to mea-sure the similarity between the preprocessed image and the confluence layer of each node of the network to determine the transfer network depth.Secondly,the feature similarity measurement was carried out on the features within the layer and between different layers,and they were sorted according to the measurement results.Again,the network was compressed by the feature metric results and the desired compression factor.The experimental results showed that using feature similarity measure to prune the transfer network channel can effectively compress the network with a small performance degradation.Therefore,using the peak signal-to-noise ratio method to measure the feature similarity can become the judgment basis for channel pruning.

关 键 词：医学图像 相似度度量 通道剪枝 网络压缩 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]