Research On Moving Object Detection And Dusty Image Restoration

Abstract

Robust principal component analysis(RPCA), a method used to decom- pose a matrix into the sum of a low-rank matrix and a sparse matrix, has been proven effective in modeling the static background of videos. However, because a dynamic background cannot be represented by a low-rank matrix, measures additional to the RPCA are required. In this thesis, we propose masked RPCA to process backgrounds containing moving textures. First- order Marcov random field (MRF) is used to generate a mask that roughly labels moving objects and backgrounds. To estimate the background, the rank minimization process is then applied with the mask multiplied. During the iteration, the background rank increases as the object mask expands, and the weight of the rank constraint term decreases, which increases the accuracy of the background. We compared the proposed method with state- of-art, end-to-end methods to demonstrate its advantages. Subsequently, we suggest novel dedusting method based on dust-optimized transmission map and deep image prior. This method consists of estimating atmospheric light and transmission in that order, which is similar to dark channel prior-based dehazing methods. However, existing atmospheric light estimating methods widely used in dehazing schemes give an overly bright estimation, which results in unrealistically dark dedusting results. To ad- dress this problem, we propose a segmentation-based method that gives new estimation in atmospheric light. Dark channel prior based transmission map with new atmospheric light gives unnatural intensity ordering and zero value at low transmission regions. Therefore, the transmission map is refined by scattering model based transformation and dark channel adaptive non-local total variation (NLTV) regularization. Parameter optimizing steps with deep image prior(DIP) gives the final dedusting result.

강건 주성분 분석은 배경 감산을 통한 동영상의 전경 추출의 방법으로 이 용되어왔으나, 동적배경은저계수행렬로표현될수없기때문에동적배경 감산에성능적한계를가지고있었다. 우리는전경과배경을구분하는일계마 르코프연쇄를도입해정적배경을나타내는항과곱하고이것을이용한새로 운형태의강건주성분분석을제안하여동적배경감산문제를해결한다. 해당 최소화문제는반복적인교차최적화를통하여해결한다. 이어서대기중의미세 먼지에의해오염된영상을복원한다. 영상분할과암흑채널가정에기반하여 깊이지도를구하고, 비국소총변동최소화를통하여정제한다. 이후깊은영상 가정에기반한영상생성기를통하여최종적으로복원된영상을구한다. 실험을 통하여제안된방법을다른방법들과비교하고질적인측면과양적인측면모 두에서우수함을확인한다.