The Cartoon-Texture Decomposition Of Image And Its Application

Object detection is an important task of image processing. Cartoon-texturedecomposition, developed recently in the field of the computational harmonic analysis,provides a powerful tool for it and has wide applications in image denoising, inpaintingand other areas. In applications of satellite optical remote sensing images, detection ofship targets plays a crucial role in monitoring and management maritime safety, searchand rescue, water transport, fisheries monitoring, illegal smuggling monitoring etc. Inthis thesis, we devoted to ship detection by using cartoon-texture decomposition ofimages. The work has important theoretical significance and application values.The classical variational based and sparse representation based models andalgorithms of cartoon-texture decomposition of images are reviewed first. Then thestatus of detection of ship targets in satellite visible remote sensing images is introduced.Through analysis of satellite visible images and their cartoon-texture decompositions,we draw some conclusions. The remote sensing ocean images may typically containships, sea, sea shore, land, islands, or clouds if there exist. Ships are generally smaller insize than other targets so that they can be efficiently decomposed into the texturecomponent while the main structure of other targets are separated into the cartooncomponent by applying cartoon texture decomposition on the observed images.Although the texture component consists mainly of ship targets, there exist some visiblesmall-scale details of the other objects which bring disturbs in detection of ships in thetexture components. Based on these observations, we propose a new ship detectionalgorithm by using the cartoon-texture decomposition. First, an image is decomposedinto a cartoon component and a texture component. Then the details of other objects inthe texture component are removed by using their priori information in the cartooncomponent. Finally, the ship targets in the texture component are enhanced by usingGabor filtering. Experimental results and data analysis show that the new algorithmefficiently improves the detection rate and significantly reduces the false rate.