Single Image Specular Highlight Removal Research On Natural Scenes

In real scenes,there are two components of specular reflection and diffuse reflection on objects exposed to natural light.When the specular reflection is relatively strong,the object surface will show the phenomenon of highlight,and it will produce saturated pixels on digital images in serious cases.Highlights bring negative effects in the task of computer vision,create discontinuity and reduce the contrast of the image,or even completely destroy the original color and detail information of the scene.Therefore,the study of image highlight removal is essential for the successful implementation of advanced visual tasks.Highlights always present different distribution in different scenarios.Previous methods of highlight removal only addressed images taken in specific background and illumination environments,reducing the influence of shooting angle and other environmental interference factors through the manual intervention.However,these methods do not conform to the features of natural scenes and can not solve the problem of saturated pixels caused by highlights and thus,introduce artifacts to images.Therefore,the thesis focuses on the problems that are difficult to be solved by existing methods based on scene feature analysis of highlight images,mainly solving the problems of restoring image information under saturated pixels caused by local directional sparse highlights and global highlight removal on natural scenes.In this paper,we first deal with the local directional sparse highlights with saturated pixels,we detect the highlight areas and extract the highlight alpha matte using the closed-form matting method based on the separation of foreground and background.Second,we prove the linear relationship between the highlight pixel values and the pixel values under normal lighting experimentally and theoretically.Starting with the image formation equation,we derive the illumination recovering operator combined with the highlight detection coefficient matrix to restore the original information destroyed by highlights in the scene.Then,we improve the general quality of highlight removal through the boundary processing techniques.In addition,this paper takes the natural scene as the main research object,using the illumination component to explain the differences between natural scenes and experimental scenes.Then,we propose a single natural scene image highlight removal model based on the dichromatic reflection model.The low-frequency information which truly represent the overall intensity changes of the specular reflection can be extracted from an image.We combine the low-frequency of the luminance and three color channels as the estimate of the specular reflection coefficient.Finally,the accurate estimate of specular reflection component is achieved by normalization and space transformation,thus the non-highlight image can be obtained.The experimental results show that our method can solve the problem of saturated pixels effectively.It performs better than state-of-the-art methods on natural scene images without introducing artifacts to images,which improves the overall image recovery quality.