Salient Object Detection Based On Sparse Subspace Clustering And Low Rank Representation

The goal of a salient object detection algorithm is to completely detect the whole salient object in the complex scene. Primates have a remarkable ability to interpret complex scenes in real time. However, in the field of computer vision, people pay more attention to the method that can accurately and effectively detect salient objects in the cluttered background.In the recent years, low rank matrix recovery(LRMR) technique has attracted much attention. Thus, several saliency detection methods based on LRMR have been proposed. Traditional LRMR based saliency detection methods assume the background lies in a low-dimensional subspace, while the small salient objects are considered as sparse noises or errors. Thus LRMR technique is usually employed to decompose an image into low-rank part plus sparse noises, where low rank part denotes the background and sparse noises indicate the salient regions. Generally, these methods can well detect salient objects of small sizes with simple backgrounds. In case of an image containing a salient object of large size, these methods usually produce higher saliency values at the boundaries of the object rather than generating a saliency map that uniformly highlights the whole object region. In this paper, we propose a salient object detection method based on sparse subspace clustering and low-rank representation. The proposed salient object detection method can completely detect a salient object of large size, and even for images with cluttered backgrounds, it can still obtain satisfactory saliency detection results.The main work of the dissertation is as follows:First, three existing LRMR based saliency detection methods are discussed in detail in the paper. These include the sparse coding and low rank matrix recovery based, the multi-task sparsity pursuit based, and the feature transformation and low rank matrix recovery based saliency detection methods. Such traditional LRMR based saliency detection methods assume that the sizes of salient objects are small and the features of background lie in a low-dimension subspace. Generally, they could not effectively detect salient objects of large sizes and fail for the images with cluttered backgrounds.Secondly, we present a novel salient object detection method based on the Laplacian sparse subspace clustering(LSSC) and low-rank representation(LRR). We segment the input image into many superpixels and group them into different clusters by using LSSC. Each cluster contains multiple superpixels that have similar features(e.g., colors and intensities), and may correspond to the salient object or a region in the background. Thus the salient object detection is reduced to the saliency detection of different clusters in the proposed method. And the latter is further achieved by performing the LRR on the feature matrix of each cluster. For that, a primitive saliency dictionary is first constructed based on the local-global color contrast of each superpixel, whose atoms can be divided into three groups, i.e., the ones mainly consisting of the features of:(1) potential foreground superpixels,(2) potential background superpixels, and(3) hybrid superpixels, respectively. Then a saliency measure is defined based on the LRR coefficients considering the following fact. For each cluster feature matrix, its LRR coefficients indicate the affinity between the cluster feature matrix and the atoms in the dictionary, i.e., the similarities between the cluster and the foreground(or background). Meanwhile, in order to detect the salient objects of small sizes, another saliency measure is also defined by using the sparse reconstruction error, as traditional methods do. The two saliency maps obtained by the two measures above are fused together to get the final saliency map.At last, the proposed method is implemented using the Matlab programming language with Windows 7 as an operating system. Several sets of experiments demonstrate that the proposed algorithm outperforms some current state-of-the-art methods. The proposed saliency detection method can not only completely detect a salient object of large size, but also accurately detect salient object of small size. Moreover, it can effectively work for the images with cluttered backgrounds.