Saliency Detection Via Sparse Reconstruction And Joint Label Inference In Multiple Features

Visual saliency plays a significant part in our vision field. By filtering out the redundant information, it efficiently helps dealing with the overload of information. It is thought that interesting (salient) objects always have some specific properties which make them different form their surroundings. So the theory is helpful to select the most interesting object in a scene. Recent years, the saliency detection model has arouses extensively interesting by many researchers and a lot of detection models have been created. The fast development of saliency detection makes it widely applied in many areas, such as segmentation, classification and content based image retrieval.Salient object is in general thought to be unique, rare, surprised or unpredicted. According to its properties, saliency detection methods can be classified as purely computational, biologically based, or their combination. Based on the motivation that the appearance of a salient target tends to be sparse in the entire scene, sparse representation has also been applied to saliency detection. However, existing sparse representation based methods often only highlight the boundaries of salient object rather than the whole object, especially for relatively large object.In the paper, we propose a novel saliency method. Given an image, we first accurately over-segment it into superpixels, and then group them into a few segments by a simplified spectral segmentation method. Next, we use the center-remaining method (a superpixel is regarded as the center and the remaining superpixels in other segments is regarded as dictionary) to reconstruct superpixel, i.e., for a segment, each superpixel it contains is described as the weighted combination of all the superpixels in the other segments. We average the reconstruction errors in a segment as its initial saliency. The hierarchical treatment helps to overcome the above problem which the salient object is not adequately highlighted. Finally, different from previous methods, which make use of multiple features by combining the saliency maps computed in individual features, we further refine saliency result by using a ranking-based inference model and define a multi-feature fitting potential to describe the interaction among multiple features.We implement a lot of experiments on four benchmark datasets which test our proposal method for quantitative evaluation. We compare the proposal method with fifteen state-of-the-art methods which provide the source code. Experimental results demonstrate that the proposed method obtain the excellent result both in precision and in recall when compared with other methods.