Salient Region Detection Algorithm Based On Minimum Convex Hull Of Feature Points And Contrast

An important feature of human vision is to allocate limited resources to the area with more information. Detection of salient region aims at simulating this feature and finds out the area where can more likely cause human’s interest. Detection of salient region has been applied to many computer vision tasks, such as image compression, target recognition, image indexing and so on. Therefore, detection of salient region is one of the most popular research focuses and challenges in computer vision.Biological studies have shown that there are significant differences between salient regions and other regular regions. Therefore, many researchers design algorithms based on image contrast. But some images also have background with great contrast. Most of algorithms work based on single contrast, so they can’t distinguish the background with great contrast from the main target. What’s more, the calculation speed is also the one the important requirements of detection of salient region algorithms. Based on the summary of classical algorithms, firstly this paper presents a new bottom features representation which based on the theory of image segmentation, normal distribution and Wasserstein distance.Then this paper apply the new representation to the classic GBVS algorithm. Secondly base on the performance of the new representation, this paper presents a new salient region detection algorithm. This new algorithm includes more principles of salient region detection. It bases on contrast and minimum convex hull of feature points. With more detection principles the new algorithm performs better than the algorithms with single detection principle. Finally this paper simulates the algorithm by computer software. Following are the three main innovations:① This paper takes superpixel as the basic processing unit. In order to build mathematic model this paper use the normal distributions of color and results of Gabor filtering to present the pixels in each superpixels. The differences between superpixels are measured by Wasserstein distance. Finally by the differences between superpixels this paper calculates the global and local contrast saliency maps.② This paper first uses the contrasts to cluster superpixels and calculates the global salient map by the responsibilities of cluster results. In order to improve the algorithms this paper uses classic center-surround operator to calculate the local salient map.③ This paper calculates the minimum convex hull of feature points and set the saliency of the region where is outside of the minimum convex is zero. Based on the distances between superpixels and the central point of convex hull, the center saliency map can be achieved. The final saliency map is consisted of the contrast saliency map and the center saliency map.This paper use Matlab simulation software and VLFeat toolkit to simulate the algorithm. Simulation experiments on the MSRA image databases were conducted. Finally the precision, recall and f-measure are calculated to evaluate this algorithm. And results of the simulation experiments show that this salient region detection algorithm performs better than some classic algorithms.