关键词: 建筑物变化检测, U-Net, 非对称卷积块, 注意力机制

Abstract:

A method of building change detection in remote sensing image based on improved U-Net is proposed. The U-Net model is used to classify the image, and the problem of change detection is transformed into a two classification problem. Each pixel in the image is divided into a change class or a non change class,the change detection results are obtained according to the change class pixels. In order to improve the robustness of the convolution kernel and prevent over fitting, asymmetric convolution block is proposed to replace the standard convolution operation of the feature extraction part of U-Net network. Aiming at the problem that the background of the image in the change detection data set is complex and the change of the small target is easy to be missed, the attention mechanism is introduced into U-Net to restrain the learning of the model to the non-change type pixel features, strengthen the learning of the change type features, and extract the more suitable features. The experimental results show that after the introduction of asymmetric convolution block and attention mechanism, the F1 score of change detection is significantly improved.

Key words: building change detection, U-Net, asymmetric convolution block, attention mechanism