Generalized Change Detection to Cue Regions of Interest

Toyon proposes to research and develop algorithms for generalized salient change detection, and to incorporate these algorithms into software tools implemented on the cloud. Our approach leverages the two most promising methods from Phase I, both based on supervised learning. The first method is the entropy-based feature vector and corresponding neural network, which we will apply at a coarse search (CS) resolution. We showed this approach to be successful at quantifying the scalar amount of salient change within an image, without regard to localization. The second method is an extension of semantic segmentation. We were able to show that this approach works well to classify salient-vs-non-salient change, on a pixel-by-pixel basis. We expect this approach to be effective for confirming and localizing (CL) change within a region flagged by our CS method. Training and testing this approach successfully will inevitably require labeled data, which we will create using both real images and a synthetic data generator. The result will be a cloud-based software system that flags large coarse regions of high likelihood or high amounts of change, and then proceeds to confirm and localize those changes.