轻量级空间移位MLP用于指静脉分割

Lightweight spatial⁃shift MLP for finger vein segmentation

下载PDF

导出

摘要基于CNN和Transformer架构图像分割网络模型参数繁多、计算复杂,需要消耗大量的内存资源,这使得它们无法满足快速、有效的指静脉图像分割需求,并且在算力有限的嵌入式平台部署非常困难。因此,提出一种基于MLP的轻量级手指静脉分割算法。首先,通过不同轴向移动特征图获取信息流来捕获局部依赖性,提高局部信息提取能力;其次,使用标记MLP块对特征图进行标记和投影卷积特征;然后,在下采样和上采样之前都添加一个轻量级注意力模块来提升分割性能,在输入到MLP的同时转移输入的通道,使网络模型更专注于学习本地依赖性。在SDU-FV、HKPU和UTFVP三个公开的手指静脉数据集中进行实验,结果表明:该方法仅使用了346.949K Params、1.835G Flops和11.023M的计算复杂度,分割性能指标Dice、AUC、Acc分别达到0.515 6、0.895 9、91.68%。在三种NVIDIA嵌入式平台上,该算法的Dice和AUC指标均取得了最优性能。 The CNN and Transformer based architectures'image segmentation network models are characterized by excessive parameters,intricate computational demands,and a substantial requirement for memory resources,which makes them unable to meet the demand for fast and effective finger vein image segmentation,and difficult to be deployed in embedded platforms with limited arithmetic power.Therefore,a lightweight finger vein segmentation algorithm based on MLP is proposed.The information flow is acquired by moving the feature maps in different axes to capture local dependencies and improve local information extraction.The feature maps are labeled and the features are projectively convolved by tokenized MLP blocks.A lightweight attention module is added to improve the segmentation performance before both down-sampling and up-sampling,shifting the inputs while inputting them to the MLP,so that the network model focuses more on learning the local dependencies.The algorithm was experimented on three publicly available finger vein datasets of SDU-FV,HKPU and UTFVP.The results show that the method uses only 346.949K Params,1.835G Flops,and 11.023M computational complexity,and the segmentation performance metrics Dice,AUC,and Acc reach 0.5156,0.8959,and 91.68%,respectively.The algorithmic method achieves optimal performance for both Dice and AUC metrics on three NVIDIA embedded platforms.

作者曾军英田慧明陈宇聪顾亚谨邓森耀尹永宏尤吴杭黄国林甘俊英秦传波 ZENG Junying;TIAN Huiming;CHEN Yucong;GU Yajin;DENG Senyao;YIN Yonghong;YOU Wuhang;HUANG Guolin;GAN Junying;QIN Chuanbo(Faculty of Intelligent Manufacturing,Wuyi University,Jiangmen 529020,China)

机构地区五邑大学智能制造学部

出处《现代电子技术》北大核心 2024年第7期54-60,共7页 Modern Electronics Technique

基金国家自然科学基金项目(61771347) 广东普通高校人工智能重点领域专项(2019KZDZX1017) 广东普通高校重点领域专项(2020ZDZX3031) 广东省基础与应用基础研究基金(2021A1515011576) 2022年广东省教育厅研究生教育创新计划项目(粤教研函[2022]1号,五邑大学-江门妇幼保健院联合培养研究生示范基地)。

关键词手指静脉分割 CNN TRANSFORMER 轻量级嵌入式平台标记MLP finger vein segmentation CNN Transformer lightweight embedded platform tokenized MLP

分类号 TN911.73-34 [电子电信—通信与信息系统] TP391.41 [自动化与计算机技术—计算机应用技术]