Research Of Salient Object Detection Based On Feature Integration

Salient object detection is an important topic of computer vision, the aim is to detect main objects in the image. As a result of the evolution and development of human vision system, human eyes have the ability of capturing the most important object in the image quickly and accurately. With the development of computer technology, people expect computer to imitate human vision to detect the most important part of the image as fast as possible, so the demand of detecting saliency by computer is appeared. The research of salient object detection is to detect the main part of the image by computer based on human attention system.Saliency detection has important research significance in many kinds of fields like computer vision, image processing, image segmentation, artificial intelligence and so on. And meanwhile, it has great application significance in many fields like image retrieval, image segmentation, object detection and tracking.Recently, saliency detection has gained extensive attention of many researchers because of its research significance, and becomes hotspot. Researchers have proposed many computing algorithms to solve the problem of detection accuracy and speed. The paper firstly reviews and analyzes previous methods. Then, it studies salient object detection approach in the view of feature integration.Firstly, most traditional saliency detection methods are based on contrast feature, and some of them have achieved great success. However, due to behavior difference of different objects, single contrast feature can’t be effective in all conditions. To solve this adaptability problem, the paper proposes to integrate color distinctness feature, boundary prior feature and objectness feature to detect saliency. To be specific, firstly, the method extracts color distinctness feature in the perspective of foreground. Next, it extracts boundary prior feature in the perspective of background. Then, it integrates the two features linearly. Finally, it extracts objectness feature, and multiply integrates the feature to enhance the saliency detection result.Secondly, some recent methods use boundary prior to study saliency, they compute saliency by considering the distribution of background. Inspired by this kind of concept, similarly, the paper proposed composition prior, and integrate composition prior feature to detect salient object. Observing from images, it can easily find that images are usually formed by following with some composition rules, such as Rule of Thirds. So the paper makes such an assumption:salient object usually placed near the composition lines. And we name this assumption as composition prior. In this paper, we treat salient object detection process as a two-classification problem and use manifold ranking algorithm to compute saliency. To be specific, firstly, according to manifold ranking algorithm we segment the image into super pixels and construct a close-loop graph. Secondly, we set super pixels which near the composition lines as query nodes and extract their color feature, then compute saliency of each super pixel by manifold ranking algorithm. Thirdly, we refine the saliency computation results in the perspective of both object and background. Considering different pixels of the same region may have different behavior, we correct their saliency based on their distance to feature center. Finally, we integrate multi-scale saliency to get the final results. The method studies saliency based on image composition lines, it computes the distribution of salient object and background by using composition prior.In comparison experiments on datasets of MSRA-1000, CSSD, and ECSSD, the two methods proposed in this paper both perform well when against the state-of-the-art methods.