CDZoom: a human-like sequential zoom agent for efficient change detection in large scenes

High-resolution (HR) remote sensing images provide rich information for human activities. However, processing entire HR images is time-consuming, and many computations are meaningless for change detection tasks since objects often cluster in local regions. To alleviate the pressure of downstream detectors, previous studies introduce a regional attention process to roughly sample candidate patches, but most solutions are tailored to particular tasks and datasets. Motivated by these, we develop a novel reinforcement learning sampling framework, and train a human-like agent, named CDZoom, to locate regions of interest by simulating human zooming behaviors. To be specific, the proposed network consists of an encoder block, multiple context blocks and a decision block. It speeds up sequential sampling operations by gradually focusing the scope of observed scene and increasing the resolution. To avoid the sparse reward problem when learning complex sampling tasks, we introduce a novel training paradigm based on curriculum learning and policy distillation. The proposed CDZoom can sample multi-size patches from multi-scale scenes, and thus generalizes well to different requirements. Experiments on public change detection datasets demonstrate the effectiveness of our method. CDZoom can reduce the computational cost by over 50%, while maintaining similar detection accuracy to models which use full HR images.