Generalized Change Detection to Cue Regions of Interest

Toyon Research Corporation proposes to research and develop algorithms for generalized change detection, by leveraging and exploringexisting and proven effective traditional and deep learning methods, with a unique 3D reconstruction component. The vast majority of themassive amounts of imagery data will have small pixel level differences due to a multitude of unimportant changes: minor misregistration,changes in lighting or occlusions, variations in weather and vegetation, etc. Toyons proposed effort addresses these challenges throughapplication of a Gaussian Mixture Model (GMM) learned via expectation maximization (EM). Within the GMM framework, we will make use ofboth standard and deep learning based features, as well as 3D model height information when available. Once salient change has beendetected, we will classify it based on examples provided by the NGA and obtained during data curation for the area of interest (AOI) wheredetection occurred. Classification will be performed by features learned via Transfer Learning from one of the large and successful CNNs(VGG16, Resnet50, Inception) and those extracted from autoencoders. Classes with few examples will make use of data augmentation, and wewill explore Generative Adversarial Networks (GANs), which have promise to be an effective and innovative exemplar generator scheme.