Feature Difference Enhancement Fusion for Remote Sensing Image Change Detection

• Published in: Pattern Recognition and Computer Vision pp. 510 - 523
• Main Authors: Hu, Renjie, Pei, Gensheng, Peng, Pai, Chen, Tao, Yao, Yazhou
• Format: Book Chapter
• Language: English
• Published: Cham Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Change detection; Deep learning; Difference enhancement
• ISBN: 9783031189128; 3031189124
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-18913-5_40

Remote sensing image change detection identifies pixel-wise differences between bitemporal images. It is of great significance for geographic monitoring. However, existing approaches still lack efficiency when dealing with the change features. The most general manner is to introduce attention mechanisms in different time streams to strengthen the features and then superimpose them together to complete the fusion of the features. These methods can not effectively excavate and apply the relationship between different temporal features. To alleviate this problem, we introduce a feature difference enhancement fusion module based on pixel position offset in the time dimension (time-position offset). We will learn the offset of the pixel changes in the corresponding areas between the bitemporal features, which will be used to guide the enhancement of the difference between the change-related areas and the change-irrelated areas in a single feature map. Meanwhile, we propose a general and straightforward change detection framework composed of the basic ResNet18 as the encoder and a simple MLP structure as the decoder, instead of the complex structures like UNet or FPN. Extensive experiments on three datasets, including LEVIR-CD, LEVIR-CD+, and S2Looking datasets, demonstrate the effectiveness of our method.