Analysis of aerospace monitoring images for novelty detection on the ground using deep learning methods

Abstract

The article is devoted to the development of mathematical software in order to detect elements of novelty in aerospace images of urban, suburban, wooded and water areas. To achieve this goal, in the course of the work, traditional and modern change detection methods, popular architectures of neural networks for segmentation, post-processing and image matching algorithms were explored. The main idea of the proposed method and the algorithms implemented on its basis is the use of the post-classification approach. It is based on the estimation of the difference between a pair of binary masks obtained as a result of segmentation of analyzed multi-temporal images using neural networks for segmentation, and specialized post-processing algorithms. The approach proposes to implement change detection in the conditions of diff erences in the geometric aspects of the compared images, which most of the previously proposed methods and algorithms have not taken into account. This approach makes it possible to use the resulting binary masks for qualitative matching of a new and previously captured image, perspective correction, and their difference, respectively. In addition, unlike most previous research, it is proposed to detect not only changed objects but also determine their classes and related topographic objects such as buildings, roads, trees, and water objects. It makes it possible to use the method in photogrammetry applications. Due to the lack of objective quantitative data to assess the accuracy of detection, visual testing was carried out. As a result, it was concluded that the proposed approach works very well only for detecting large changes since the implemented detection method is highly sensitive to the smallest changes.