Mask-CDNet: A mask based pixel change detection network

Change detection between multi-temporal images becomes a core technique widely used in various fields. But there still exist some challenging issues in the field of change detection. This thesis mainly focuses on the issue that two compared images captured at different times are hard to be precisely aligned due to the fact that the camera is mounted on a moving platform. This issue will lead to the problem that it is difficult to determine the change area between a pair of roughly registered images in pixel-level. To conquer this problem, a novel deep learning based change detection framework, consisting of two collaborative modules, is proposed to improve the estimation accuracy and computation efficiency. The first module is mainly used to roughly predict change areas and match information of two unaligned compared images in the absence of ground truth. The second module aims to refine the change areas and make results more accurate and interpretative. Because of the designed two modules, the framework could not only deal with two roughly registered images, but also be robust to uninteresting change coming from noise or arbitrary spurious differences. Although it is a two-stage method, the proposed framework is trained end-to-end. Meanwhile, the design of pure network make it efficient. Experiments evaluated on PCD-2015 dataset and AICD 2012 dataset demonstrate the method outperforms existing literatures. (C) 2019 Elsevier B.V. All rights reserved.