Image Saliency And Its Evaluation

Currently, saliency analysis has been very successful in many areas of application, such as image segmentation, image object extraction and adaptive image compression and coding. Saliency analysis algorithms are two types: one is based on the low-level visual information, data-driven bottom-up attention model; one is based on high-level visual information, knowledge-driven top-down attention model. Presently, most algorithms are the former. We do not have a system of evaluation system for the current algorithms.This paper introduce two statistical indicatores to discuss the advantages and disadvantages of several typical algorithms, basing on the evaluation experiment we try to build a new saliency analysis algorithm.The first part introduce several typical significance analysis algorithms, there are Itti etal proposed saliency map model, Harel etal proposed based on graph theory image saliency analysis algorithm (GBVS algorithm), Achanta etal proposed two of the saliency map model (AC algorithm and the IG algorithm) and Xiaodi Hou made such a residual spectral method (SR method). Introducing these algorithms just because they are based on different theories, very representative, and easy to implement on the computer, the results were better experiments, is quoted a higher rate of several algorithms.In the second part of this article, we introduce two statistical indicatores. Comparing the advantages and disadvantages of five algorithms. Based on manual segmentation three groups evaluation is proposed (respectively, saliency map and manual segmentation image contrast; saliency map based on a fixed threshold segmentation map and manual segmentation image contrast; adaptive threshold segmentation map and manual segmentation image contrast). Analyze the advantages and disadvantages of these types of algorithms from experimental results, and try to analyze the main reason so to think about how to build a better new algorithm.In the third part, we try to create a new algorithm.Calculating the evaluation index compare to algorithms introduced in the first part of this article.Finally, use the application of"Seam carving for content-aware image resizing"to discuss its feasibility and adaptability.