Research Of Shadow Removal Based On Invariant Image

Shadow, as a kind of image degradation, is caused by illumination reduction.The shadow will lead to the loss of target information in imageries, and reduce theimage interpretation accuracy, which will seriously disturb image quantitativeanalysis and applications.To reduce the adverse effects of shadow, it is essential to remove shadow fromimages. Current shadow removal methods can be divided into two categories:methods based on Poisson equation and methods based on shadow factor estimation.Shadow removal based on Poisson equation recovers shadow information by settingthe gradient of shadow boundary to zero, which leads the loss of the textures inshadow boundary. In shadow removal based on shadow factor estimation, shadow isdivided into umbra and penumbra, and shadow factor is estimated respectively. Therecovered results are better than that of the former methods. However, the interactivemethod needs to select shadow boundary artificially. Furthermore, the boundarylocation of shadow is crucial for shadow removal results in these two methods. Andshadow boundary locating is a challenge task in the images with complexity scene orrichly-texture. To improve these defects of the algorithms, the research andinnovation of this paper are summarized as follows:1. The spectral and geometric properties of shadow have been analyzed. Andthe current shadow removal algorithms have been summarized. A new shadowremoval method without detecting shadow edges has been proposed, in which aninvariant image is built to restore image by fitting linear relationship between pixelvalues in invariant image and in raw image. In traditional method, the invariantimage is obtained by rotating the projection direction from0to180in step of1. The traversal algorithm is time-expensive and difficult to obtain optimal solutionfor the constraint of step precision. In this paper, Fisher discrimination criterion is applied to derive the accurate invariant direction automatically.2. A new shadow removal method based on gradient field is proposed. Firstly,shadow boundary is detected approximately. Then, the gradients in internal shadowregion and in shadow boundary are modified respectively to obtaining thenon-shadowed gradient field. Based on the gradient field, the information in shadowregions is recovered with Poisson equation. This method can preserve the textures inthe shadow regions and isn’t sensitive to the accuracy of shadow boundary.Through the above work, shadow in the images can be removed effectively;meanwhile the textures are well recovered. Several images have been chosen to testthe feasibility and effectiveness of the proposed method.