LR Aerial Imagery Categorization by Transferring Cross-Resolution Perceptual Experiences

Hundreds of satellites orbiting at various altitudes capture an extensive array of aerial photographs daily. High-altitude satellites typically acquire low-resolution (LR) images that cover vast areas, whereas their low-altitude counterparts obtain high-resolution (HR) images detailing much smaller regions. The accurate interpretation of LR aerial imagery is crucial in the field of computer vision, yet it presents significant challenges, including the complexity of emulating human hierarchical visual perception and the daunting task of annotating enough data for effective training. To address these challenges, we introduce a cross-resolution perceptual knowledge propagation (CPKP) framework, which aims to leverage the visual perceptual insights gained from HR aerial imagery to enhance the categorization of LR images. This approach involves a novel low-rank model that segments each LR aerial photo into distinct visually and semantically significant foreground regions, alongside less pertinent background areas. This model is capable of generating a gaze shifting path (GSP) that reflects human gaze patterns and formulating a deep feature for each GSP. Subsequently, a kernel-induced feature selection algorithm is deployed to extract a concise yet powerful set of deep GSP features that are effective across both LR and HR aerial images. Utilizing these features, a linear classifier is collaboratively trained using labels from both LR and HR images, facilitating the categorization of LR images. Notably, the CPKP framework enhances the efficiency of training the linear classifier, given that HR photo labels are more readily available. Our comprehensive visualizations and comparative analysis underscore the effectiveness and superiority of this innovative approach.