Abstract: This paper proposes a lightweight scene-level change detection network MIDF-Net, designed to address the issues of high background noise impact and low detection efficiency in large-scale building change detection within complex airport clearance zones. MIDF-Net consists of three parts: a dense connection feature extractor, a differential feature extractor, and a multi-instance classifier. The dense connection feature extractor uses a siamese dense connection network to extract dual temporal image features, while the difference feature extractor focuses on generating variation features by combining dual temporal image features. The multi-instance classifier obtains scene classification results from key local semantic features. Using image data from airport clearance zones across seven different cities, this study constructs a building change detection dataset. Experimental results show that MIDF-Net achieves high detection performance on this dataset. Furthermore, ablation experiments verify the efficacy of each module in MIDF-Net.

Key words: airport clearance zones, change detection, deep learning, dense connection, differential feature extraction, multiple instance learning, multiple instance and differential feature net (MIDF-Net)