Salient Region Detection Based On Color Contrast And Anisotropic Diffusion Segmentation

Saliency detection has become a valuable tool in computer vision processing. The pre-processing of high level application such as adaptive compression, image retrieval, and object recognition can be accomplished by saliency detection. In this paper we reconsider some designs choices of previous method and propose a superpixel-level and region-level combined saliency detection. Our saliency detection method takes both global color dissimilarity and spatial consistency into consideration.Our algorithm consists of four basic steps. First, we get a coarse salient region by detecting interesting points. The coarse salient region is characterized by containing most salient object and part background and the image except this region basically composes of background only. Two saliency maps separately considering details and integrity are achieved by color contrast and spatial relation measure based on the coarse salient region. Every image is decomposed into superpixels representing mid-level visual information. We rank the saliency value of each superpixle by the color dissimilarity with the coarse background, then we have superpixel-level color contrast saliency map. The heat anisotropic diffusion equation solves the image region division. The region division is modeled by temperature maximization on anisotropic heat diffusion, of which the temperature maximization with finite K heat sources corresponds to a K-way region division that maximizes the region division confidence of every pixel in an image. Combine color contrast saliency map as one dimension with tliree dimension color features into region division separate the salient object from background in the most extent. We rank the saliency value of each region by overlap with the coarse salient region, then we have region-level position relation saliency map.We obtain the final saliency map by multiplying color contrast saliency map by position relation saliency map, which is characterized by highlighting the salient region and glooming the background resulting a separate, compact and consistent salient object standing out the image. Our algorithm consistently outperformed six representative saliency detection methods, yielding higher precision and better recall rates, when evaluated using one of the largest publicly available data sets-microsoft database.