Salient Object Segmentation Model

Visual perceptions play an irreplaceable role when humans acquire knowledge from their surroundings, and the image is one of the most important carriers of visual per-ception. With the wide popularity of smart phones, image resources on the internet are explosively growing. If it is possible to simulate human visual perception mechanism, and automatically find out the most important and significant part of the scene, this will help to break through the bottleneck of information processing, saving resources of calculation and analysis and improving the speed of processing. Salient object is the most impor-tant and attractive region of a scene. Salient object segmentation aims to automatically and accurately extract the most salient object from an input image. In recent years, salient object segmentation model has attracted much interest from researchers in the computer vision. And it has been applied in the robot vision, video security surveillance, object detection and recognition, video compression and coding, motion detection, visual tracking, etc. This work starts from the region based image segmentation model, and steps into the salient object problem. Incorporating with the subspace clustering theory, the anisotropic diffusion equation, flow diffusion related theory, some novel models have been proposed. At last, it has been briefly introduced about the application of the salient object segmentation in the iNavigation project. This work mainly includes the following parts:1. In the study of region based image segmentation, a new non-convex half norm based sparse and low rank coupling image segmentation model has been presented. Sub-space clustering model has been introduced into the image segmentation problem. S1/2 norm is used as the low rank representation instead of the nuclear norm, and L1/2 is used as the sparse representation instead of the L1 norm. It has been demonstrated that the S1/2 norm and L1/2 norm are the better approximations for the low rank representa-tion and the sparse representation, respectively, although they are both non-convex. In addition, their minimum problems both have closed solutions. The variables are sepa-rated through the ALM method, and an effective algorithm via the half norm operator is proposed. Amount of experiments demonstrate the accuracy of segmentation and the robustness to the outliers.2. We introduce the nonconvex Schatten-q regularizer for the subspace clustering problem in order to solve the rank minimization problem. In this context, we present the GMST algorithm, a new generalized matrix soft thresholding algorithm, to solve the Schatten q regularizer minimization problem. The proposed method always obtains a solution with a lower rank than the other methods. This shows that the GMST algorithm has the ability to depict the structure of the redundant data to a much greater extent than the existing methods. A large number of experiments demonstrate that the proposed method is competitive to the state-of-the-art methods, but has a lower computational cost and is especially more robust to outliers. Further more, our newly proposed solver to Schatten-q (0< q< 1) regularizer is more accurate. Many current solvers to Schatten-q regularizer have reported that when q= 1, their methods will become the widely used singular value thresholding algorithm. Beyond that, when q= 0.5, our newly proposed solver also coincides with S1/2 regularizer based half thresholding algorithm. A rigorous mathematical proof is given.3. A nonlocal anisotropic diffusion equation is constructed to model the evolution of visual saliency. A new multi-directions discretization scheme is adopted to solve the equation. Visual attention diffusion is modeled as a series of diffusion progresses until achieving a stable status. So two stages of diffusion are involved. In the first stage, image boundaries are used to guide the background iterative diffusion process, and the initial saliency seeds are obtained by sorting. In the second stage, an optimization of saliency seeds is executed, and the good saliency seeds are strengthened and updated during each iteration. The final saliency map is obtained through iteratively diffusing saliency scores from the optimal saliency seeds (i.e. the most representative salient elements). Extensive experimental results on two large benchmark databases demonstrate the effectiveness of the proposed method.4. Visual attention spreading is formulated as a nonlocal diffusion equation. Differ-ent from the other diffusion based methods, a nonlocal diffusion tensor is introduced to consider both the diffusion strength and the diffusion direction. With the help of diffusion tensor, along with the principle direction, the diffusion has been suppressed to preserve the dissimilarity between the foreground and background, while in other directions, the diffusion has been boosted to combine the similar regions and highlight the salient object as a whole. Through a two-stages diffusion, the final saliency maps are obtained. Exten-sive quantitative or visual comparisons are performed on three widely used benchmark datasets, i.e. MSRA-ASD, MSRA-B and PASCAL-1500 datasets. Experimental results demonstrate the superior performance of our method.5. In the current state-of-the-art literatures on salient object detection, the focus is in finding one or several more discriminative features to segment the salient object from the background. However, in the analysis of complex scenes, most techniques are chal-lenged by noise, granularity and regions in the scene image with similar pixel intensities. Inspired by the feature integration theory in cognitive psychology, it is noticed that the salient objects can be associated with the image regions that are consistently distinct in most of the feature spaces. Base on this point, the feature distinctions are computed in each feature space respectively, and a saliency flow model is proposed to formulate the process of the saliency spread directly. Both low level and mid-level features are involved. Finally, the saliency map is obtained through fusing the feature distinction maps with the tuned weights after a post-processing. The consistent feature distinctions are free from the specific elaborate features and represent higher robustness in the complex scenes. This also benefits our model. Extensive experiments on six public benchmark databases demonstrate the robustness and the superior performance of the proposed method.