Cross-Resolution Perceptual Knowledge Propagation for Low-Resolution Aerial Photograph Categorization

The ability to precisely identify the semantic categories of low-resolution (LR) aerial images is a critical skill in the field of remote sensing. However, accurately performing this task is challenging due to several factors: 1) the complexity of incorporating human visual perception into the recognition algorithm; 2) the significant human effort required to annotate a large number of training LR aerial images; and 3) the difficulty in selecting high-quality features for accurate categorization. To address these challenges, we introduce a groundbreaking approach known as cross-resolution perceptual knowledge propagation (CPKP), which aims to utilize the visual perceptual insights gained from high-resolution (HR) aerial images to improve the categorization of LR images. In particular, our approach draws inspiration from the human visual system to create a novel low-rank model. This model is designed to separate each LR aerial image into multiple visually and semantically prominent foreground regions alongside less noticeable background areas. It accomplishes two key objectives: 1) generating a gaze shifting path (GSP) that replicates the sequence of human gaze shifts and 2) computing the deep features for each GSP. Subsequently, we develop a kernel-induced feature selection (FS) algorithm to extract a concise, yet effective set of deep GSP features that can distinguish between LR and HR aerial images. Utilizing the labels from both LR and HR images, we then collaboratively train a linear classifier for the categorization of LR images. Notably, our CPKP framework significantly enhances the efficiency of training the linear classifier, mainly because acquiring labels for HR aerial photographs is typically more feasible than for LR images. The effectiveness of our method is further demonstrated through extensive visual results and a comparative analysis, showcasing its superiority.