改进CycleGAN的半监督建筑物提取算法  

Improved CycleGAN Algorithm for Semi-Supervised Building Extraction

作　　者：卢鹏[1] 仲闯 LU Peng;ZHONG Chuang(College of Information,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306

出　　处：《计算机工程》2025年第3期241-251,共11页Computer Engineering

基　　金：上海市科技创新计划项目(20dz1203800);地方高校能力建设项目(19050502100)。

摘　　要：建筑物提取需要大量的标注数据进行训练,收集和标注数据需要耗费大量时间。为了在小样本遥感图像数据集上基于半监督学习实现建筑物提取的目的,构建4组建筑物提取数据集,提出了一种基于循环一致性生成对抗网络(CycleGAN)的建筑物提取算法。首先,在生成器中引入全局注意力机制(GAM)以增强对建筑物和图像背景细节特征的区分;其次,在判别器中加入谱归一化层以增强训练稳定性,解决了训练过程中梯度消失问题;最后,改进对抗损失和循环一致性损失以提高生成图像的质量,避免生成图像的过度平滑化,并引入Identity损失以限制生成器不会自主修改输入图像的颜色,保证输入图像与输出图像颜色组成的一致性。实验结果表明,在第1组小样本数据集上,与UNIT、MUNIT、U-GAT-IT、SPatchGAN、QS-Attn模型进行半监督实验对比,结构相似性(SSIM)值和准确率分别至少提高了3、8.1百分点,在扩充数据规模的数据集上,使用改进后的算法进行全监督和半监督实验对比,验证了改进后的算法在小样本遥感图像数据集上实现建筑物半监督提取的有效性。Currently,building extraction requires a large amount of labeled data for training,and it takes a long time to collect and label the data.In order to achieve building extraction based on semi-supervised learning on a small sample remote sensing image datasets,this study constructs four sets of building extraction datasets and proposes a building extraction method based on a Cycle-consistency Generative Adversarial Network(CycleGAN).First,this study introduces a Global Attention Mechanism(GAM)module into the generator to enhance the distinction between the details of the buildings and image backgrounds.Second,it adds a spectral normalization layer to the discriminator to enhance the stability of the training,which solves the problem of gradient vanishing during training.Finally,it improves adversarial and cycle consistency losses to improve the quality of generated images,avoids excessive smoothing of generated images,and introduces Identity loss to limit the generator from modifying the colors of the input images involuntarily to ensure consistency of the color composition of the input and output images.The experimental results show that on the first small-sample dataset,the Structure Similarity Index Measure(SSIM)and accuracy increase by at least 3 and 8.1 percentage points,respectively,compared with those of the UNIT,MUNIT,U-GAT-IT,SPatchGAN,and QS-Attn models.On a dataset with an expanded data scale,the improved algorithm is used to compare fully supervised and semi-supervised experiments,and the effectiveness of the improved algorithm in realizing the semi-supervised extraction of buildings on a small-sample remote sensing image dataset is verified.

关 键 词：建筑物提取 循环一致性生成对抗网络 谱归一化 全局注意力机制 半监督 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]