Image Saliency Extraction And Application Based On Small Regions

With the development of science and technology, images become importantsources of information in many fields. And to process images efficiently andaccurately become popular research directions. Human visual system is able toquickly focus on a few salient objects in face of a complex scene. A great deal ofimage processing applications is based on some key image components, therefore, toextract salient objects or regions automatically will do great favor to many imageapplications. Saliency extraction is such a process as to extract saliency regionsaccurately in spite of redundant information. In this paper, we propose a simplealgorithm to detect saliency regions from images which is more consistent withhuman vision system.Furthermore, we apply our saliency extraction method to someother image applications such as image segmentation and image retrieval.The key idea of many image segmentation algorithms is to find a suitablethreshold,which is able to detect key objects and ignore background areas. In thispaper, we combine our saliency extraction algorithm with image segmentation processto improve its performance. On the other hand, content-based image retrieval methodis widely used and its main principle is to extract low level features such as color,texture, shape and etcto do image matching. In this paper we try to improve imageretrieval performance by matching salient parts of images.Overall, we have completed the following tasks in this paper:(1) In this paper, we propose a novel image saliency region extraction method.We find that most of exist image saliency extraction methods are based on pixels.However, this kind of methods is easily affected by noise pixels. We provide a simpleapproach to overcome this shortcoming. In this paper we firstly divide a n image intoregular small regions and then extract important image features such as color contrast, relative position and pixel complexity to obtain the final extraction method. Wecompare our method with other five state-of-the-art saliency extraction models in ourexperiments.(2) Another standard to judge the performance of saliency extraction method isimage segmentation application. We analyze some classic categories of segmentationalgorithms and find out their advantages and disadvantages. In this paper, we designtwo types of segmentation method: fixed threshold segmentation method and adaptivethreshold method. We compare our segmentation methods with other segmentationresults obtained by five other state-of-the-art saliency extraction models using p-rcurve, average recall, average precision and F-measure as evaluation factors.(3)In this paper we apply our image saliency extraction method to improveimage retrieval process. We focus on content-based image retrieval methods whichusually employ some key areas to reduce the amount of calculation. So our idea is touse the image saliency extraction method to refine the feature point and simplify thealgorithm. We improve the image retrieval algorithm based on generalizedco-occurrence matrix. In the end of this paper, we compare two image retrieval methodsand use actual retrieval results and precision histogram to evaluate the retrievalperformance.