Aerial Photograph Categorization by Cross-resolution Deep Human Gaze Behavior Learning

Accurately recognizing aerial photographs is a useful technique in many domains like autonomous driving and environmental evaluation. In practice, both low-resolution high-resolution photos are captured asynchronistically for each region, as there hundreds of earth observation satellites orbitting the earth. Realizing such multi-resolution-based region recognition difficult task due to three challenges: 1) mimicking human visual perception when they actively viewing semantic objects inside photo; 2) deeply modeling visually/semantically salient sequentially perceived by system; 3) developing cross-resolution knowledge transferal module enhance feature representation an area. To solve these challenges, we propose cross-domain photograph system leveraging spatial composition deep encoding gaze shifting path (GSP) with high-resolution. More specifically, first use active learning algorithm discover multiple object patches constructing GSP from photo. Then, aggregation-based model formulated link features learned GSP. Subsequently, novel leverages global counterparts upgrade deeply-learned Using upgraded feature, multi-label SVM classifier trained categorizing photographs. Comparative studies on our million-scale set have demonstrated competitiveness approach.