Saliency Detection Based On Superpixel Segmentation

The primary function of visual system is to interpret the surroundings of people’s life and exchanges information to occurrence.Human visual has a capability of searching for the target quickly, that is saliency detection, but it is a challenging technology for the computer vision. This technology can help to find and extract the part that is important,usefuland noteworthy from the massive data.The existing detection models aremostly used for simple backgroundwhile the detection results of image with complex background are not accurate which increase the complexity of the subsequent image processing.How to detectand segment out the significant object is the main work of this paper.In this paper, a novel visual saliency detection model based on superpixel segmentation is proposed. This model can filter out the salient region from natural scenes automatically, especially images with complex background. The main work is as follows:(1) Adopt the superpixels segmentation during the pre-processing. Cluster the similar pixels based on the features of color, shape, edge and luminance, and divide the image into several irregular areas by the superpixels segmentation. The subsequent processing is conducted on the image blocks, which greatly reduces the time complexity.(2) Propose LBP-like method to extract background template. The four boundaries are re-expressed by local binary encoding based on the color, luminance, and texture features, and the boundary with the greatest value is discarded. Using the geodesic theory, this paper combines HLBP algorithm with LAB color components to extract texture features and calculates the shortest path from each segment block to the background template. By this way, the coarse saliency map can be obtained. Propose a post-processing algorithm, the OTSU segmentation algorithm and the Object-biased Gaussian model are used to highlight the foreground while suppressing the noise of background to get the accurate saliency map.(3) The saliency map is applied to the image compression and the results are compared with the JPEG 2000 ordinary algorithm.This paper takes the simulation on two image databases: the MSRA database which contains 1000 images and the BSD database which contains 300 images with complex background to verify the validity of the proposed model. The proposed model is compared with the seven state-of-the-art models, and the subjective evaluation and objective evaluation prove the effectiveness of the algorithm. Besides,the saliency model is applied to the ROI encoding based on JPEG 2000 for image encoding, and experimental results prove the saliency map can be applied to the JPEG 2000-ROI compression.