面向医学和遥感图像的基于GhostNet的轻量级U-Net改进研究和对比  

Improvement and Comparison of Lightweight U-Net Based on GhostNet for Medical Imaging and Remote Sensing Images

作　　者：郑艺 ZHENG Yi(The 27th Research Institute of China Electronics Technology Group Corporation,Zhengzhou 450047,China)

机构地区：[1]中国电子科技集团公司第二十七研究所,郑州450047

出　　处：《电光系统》2024年第4期13-17,共5页Electronic and Electro-optical Systems

摘　　要：U-Net广泛用于医学影像学和遥感图像的语义分割。针对语义分割模型庞大,极其占用CPU运行资源,运行速度慢等问题,通过U-Net轻量化改进,可达到快速识别的效果。通过轻量级特征提取网络GhostNet系列的廉价操作对U-Net轻量化改进,为U-Net的轻量化研究提供参考。目前,ChostNet系列拥有GhostNet,G-GhostNet和GhostNetV2三个版本,它们各有优势。首先,依次将GhostNet,G-GhostNet和GhostNetV2的主干部分作为U-Net的编码器嵌入模型。为实现编码器与解码器匹配,用网络的单元模块代替U-Net的卷积运算。利用医学公开数据集和遥感图像数据集对三种模型进行训练、验证和测试,获得模型的各项性能评分。最后对比三种模型的各项性能评分。实验结果表明:遥感图像分割任务中,基于G-GhostNet的轻量化模型效率最高,在损失少量精度的情况下获得非常快的速度,实现实时分割。而医学影像学分割任务中,基于GhostNet的轻量化模型速度较快,且分割精度远高于基于G-ChostNet的模型,模型的效率更高。GhostNetV2在两种分割任务中均不占优势。U-Net is widely used in semantic segmentation of medical imaging and remote sensing images.Due to the large module size,extremely high GPU resource and slow running speed,U-Net lightweight is improved to achieve fast recognition.By utilizing low-cost operations of lightweight feature extraction network GhostNet series,the lightweight improvement of U-Net is made,providing a reference for lightweight research of U-Net.At present,the GhostNet series has three versions:GhostNet,G-GhostNet and ChostNetV2,and have their respective advantages.Firstly,the backbone of them is sequentially embedded in model as encoder of U-Net.To achieve encoder-decoder matching,the convolution of the decoder is replaced by the unit of the GhostNet series.Three modules are trained,validated and tested with public medical datasets and remote sensing image datasets to obtain various performances scores.Finally,the performance scores of the three modules are compared.The experimental results show that the lightweight module based on G-ChostNet has the highest efficiency in remote sensing image segmentation tasks,achieving very fast speed with minimal loss of accuracy and obtains real-time segmentation.In medical imaging segmentation tasks,the lightweight module based on GhostNet has a faster speed and much higher segmentation accuracy than the module based on G-GhostNet,resulting in higher efficiency of the module.GhostNetV2 does not have any advantages in both segmentation tasks.

关 键 词：轻量级网络 U-Net ChostNet 分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] R319[自动化与计算机技术—计算机科学与技术]