基于卷积神经网络的图像分割方法研究  

Research on Image Segmentation Methods Based on Convolutional Neural Networks

作　　者：戚伟[1] 葛斌[2] 桑冬青[1] QI Wei;GE Bin;SANG Dongqing(School of Intelligent and Electrical Engineering,Huainan Vocational and Technical College,Huainan Anhui 232001,China;School of Computer Science and Engineering,Anhui Universityof Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]淮南职业技术学院智能与电气工程学院,安徽淮南232001 [2]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《长春工程学院学报（自然科学版）》2024年第1期85-89,共5页Journal of Changchun Institute of Technology：Natural Sciences Edition

基　　金：安徽省高等学校自然科学研究重点项目(KJ2020A1163)。

摘　　要：针对传统卷积神经网络存在的参数多、过度拟合导致图像分割精度不高、算法运行效率低的问题,采用最大池化处理取代下采样层,构建改进的CNN结构,获得U-Net卷积神经网络,并进行进一步改进。将改进的U-Net卷积神经网络应用于高分辨率的遥感图像中,结果表明其可以对遥感图像中的小建筑物进行精细、完整分割。另外,通过和FCN32s、SegNet、FCN8s的对比,指出改进的U-Net卷积神经网络在遥感图像分割中具有更加良好的性能。Aiming at the problems of multiple parameters,overfitting leading to low image segmentation accuracy and low algorithm efficiency in traditional convolutional neural networks,maximum pooling processing is adopted to replace the downsampling layer,and an improved CNN structure is constructed to obtain the U-Net convolutional neural network,which is further improved.The improved U-Net convolutional neural network is applied to high-resolution remote sensing images,and the results show that it can perform fine and complete segmentation of small buildings in remote sensing images.In addition,by comparing with FCN32s,SegNet,and FCN8s,it is pointed out that the improved U-Net convolutional neural network has better performance in remote sensing image segmentation.

关 键 词：卷积神经网络 图像分割 遥感图像 

分 类 号：TP391.2[自动化与计算机技术—计算机应用技术]