Intrinsic Image Decomposition:Algorithm And Application

First, a hierarchical approach is proposed to single image intrinsic decomposition based on L0 sparse representation. Compared to previous methods which only built local reflectance constraints, the proposed method can build sparse correlation of reflectance between non-local pixels which enables the proposed method to produce globally consistent reflectance. Besides, the hierarchical framework promotes the efficiency of the proposed approach and makes the proposed method much less dependent on chromaticity, hence more robust to natural images. Homogenous smoothness prior, brightness scale of shading and the sparse constraints on reflectance together construct the final intrinsic image decomposition model. The decomposition model can finally be formulated as a quadratic minimization problem, which can be efficiently solved in a closed form. Visually and quantitatively comparison to state-of-the-art methods on standard dataset and natural images validates the efficiency of the proposed algorithm.Second, a new problem called co-intrinsic images decomposition is proposed,that performs intrinsic decomposition on multiple images, which share the same foreground with arbitrarily different illuminations and backgrounds. The common foreground across different images is demanded to have same reflectance values after cointrinsic decomposition. For the purpose of efficiency, superpixels are used to represent reflectance, which greatly reduce the number of unknowns. A uniform approach is utilized to automatically derive non-local reflectance relationships via L0 sparse representation between superpixels from intra- and inter-images. Based on a unicolor-light image model, homogenous smoothness prior, brightness scale of shading and the sparse constraints on reflectance together construct the final co-intrinsic images decomposition model. Extensive experiments show plausible results of proposed approach in preserving consistent reflectance of common foreground in multiple images. The benefits of the proposed method are also validated in boosting the accuracy of image co-saliency detection and rectifying illumination differences in small change detection.