Change Detection of Non-Fixed Targets in Low-Altitude Unmanned Aerial Vehicle Inspections Based on Style Transfer

Abstract

With the rapid development of UAV technology, the demand for detecting changes in targets during low-altitude inspections is increasing. In low-altitude inspection scenarios, natural changes account for a much larger proportion than unnatural changes. Unsupervised change detection based on statistical and clustering algorithms often results in false detections of the more prevalent natural changes, leading to decreased detection accuracy. To address this issue, this paper proposes a low-altitude inspection change detection model (LPCD) based on style transfer. The model extracts features through an encoder and uses differential attention to analyze style differences. An adaptive instance normalization (AdaIN) module in the decoder ensures natural style consistency. Reconstruction loss between generated and source images in unnatural change regions is used with mapping and thresholding to improve the detection of unnatural changes. Compared to existing change detection algorithms in the remote sensing domain, the proposed model achieves improvements in accuracy of 0.01 and 0.01 on two data sets, respectively. F1 scores increase by 0.14 and 0.3, and the false alarm rate is reduced to 0.025 and 0.021.