Image Saliency Detection Based On Superpixel Segmentation And Merging

Image saliency detection is a very important part in many computer vision applications.Saliency describes the differences between the surrounding background and the target.The main purpose of saliency detection is to calculate the area in the image that the salient target may occur and filter out the redundant background information from the image.This can arrange the computing resources more efficiently and improve the computational efficiency.The value in saliency map at a pixel in image shows the possibility whether the pixel is belong to the salient target.The value expressed the size of the corresponding pixel in the image by the degree of attention.The larger value of the pixel the more probably it belongs to the final salient target.We propose an image saliency detection algorithm based on the superpixel segmentation and merging in this paper.Different from other bottom-up image saliency detection algorithms,we first build a more accurate convex hull to find the location of the salient target.Then,we build the initial saliency map in different way.Finally,we use the accurate convex hull to update the initial saliency maps under the Bayesian framework and merge the different update result to build the final result.Main innovation are as follows.First,many algorithms use interest point to construct convex hull or use the sliding window to locate the target.These methods always contain the problem of location inaccuracy or the background is included.In this paper,we construct the more accurate convex hull by segmenting and merging superpixels.This paper combines a variety of image segmentation methods to get the convex hull.These methods include level set method,SLIC(simple linear iterative clustering),construction of convex hull with interest point,fuzzy image to construct convex hull.The different segmentation results are optimized by edge detection and we propose a small superpixel processing methods.Second,for the initial saliency map,this paper constructs a saliency map based on color histogram to reflect the global statistical information of different colors in the image.Then a saliency map based on the regional contrast is constructed to reflect the spatial position and color feature information between different regions in the image.Third,a new fusion algorithm of update saliency maps based on Bayesian framework is proposed.Using the Bayesian framework to update the saliency map based on color histograms and region contrast with the use of accurate convex hull.Then the two update saliency maps are merged to obtain the final saliency map.We experimented on four open standard databases and compare the result with the corresponding ground truth with artificial marks.The results were evaluated on a variety of different evaluation criteria.Compared with other algorithms,the saliency detection algorithm proposed in this paper can get better results.