Abstract: Accurate building segmentation in remote sensing images remains a significant challenge due to varying building sizes, occlusion by trees and unstable illumination. The convolutional neural network (CNN) model often loses high-frequency details such as target boundaries and fine structures. This makes the precise segmentation of buildings in remote sensing images a challenging problem. To solve this problem, this paper proposes a deep convolutional neural network model based on contour measurement. By introducing the Sobel edge detector, the network obtains additional edges to enhance the boundary of image segmentation in an unsupervised manner. In addition, a denoising module is incor-porated to suppress noise hidden in low-level features. During training, in addition to the commonly used Dice coefficient and cross-entropy loss, a contour constraint loss function is introduced to further enhance the edge information and preserve the geometric topology of the buildings. This method is tested on the remote sensing images of buildings from the Inria Aerial Image Labeling dataset and Massachusetts Buildings dataset. Experimental results show that the proposed model effectively captures the edge details of weak light and occluded targets, thereby improving the accuracy of building segmentation. The proposed model achieves an average intersection over union (IoU) of 0.7860 and 0.7655, and a boundary IoU of 0.7359 and 0.7168, respectively, indicating enhanced accuracy in both regional and edge-level evaluation.

Key words: remote sensing image, contour constraint, contour enhancement, feature denoising, convolution neural network