Research On Image Salient Region Detection Methods And Applications

A profound challenge in computer vision is to make the computer understand the sur-rounding scene via image or video. While three vital tasks of this field are focusing on the key object, detecting the shape and contours, and capturing the context information. To achieve this goal, obtaining the visual salience information from the digital image is the most basic step. The visual salience information can significantly reflect the de-gree of stimulation to the human visual system from different regions of the image. The image salient region detection technology can be used to extract region salience infor-mation from the image. Reliable saliency estimation breaks down the barrier between the content understanding and underlying characteristics, making the higher level under-standing of the image possible; what is more, it furnishes valuable reference information for various applications including content-aware image editing, image segmentation and image retrieval. Exhaustive study of saliency detection has a far-reaching significance in improving the performance of image analysis and image understanding systems, as well as in enhancing the level of application of image processing technologies.The salient region detection method produces salience value of each region in an image, following the visual saliency laws and utilizing a variety of basic image features comprehensively, such as color, intensity, local orientation, etc. Many salient region detection methods have been proposed so far, but there are still some challenges to solve this problem:1) The quality of the saliency detection results is not high. The saliency map pro-duced by the detection methods cannot reflect the saliency information of the image correctly and completely. They are susceptible to confused color, complex texture of the object and changing background in the image. The saliency map cannot emphasize the most salient object, highlight whole salient regions uniformly and establish well-defined boundaries of the salient object. So the saliency map cannot be used in some application stably and befittingly.2) The ways and means of the use of saliency information is still relatively primary. Although saliency information is used to try solving some problems, it cannot obtain satisfactory results in many applications. So the potential of saliency information cannot be played fully. More valuable applications need to be found to extend application scenarios of saliency information.To solve these problems, the research of this thesis are from following aspects:1) Saliency detection process is modeled as two stages:a primary stage and an ad-vanced stage. In the primary stage, the basic saliency of each region is detected. And in the advanced stage, the most salient region is selected. Its saliency value is strengthened and the saliency value of its surrounding area is weakened.2) After analyzing a large number of state-of-the-art detection methods, two good ideas of designing saliency detection methods are summarized. One idea is to use frequency-tuned detection technology and the other is to use region-based global con-trast technology. Along these ideas, the new methods, weighted multi-scale frequency-tuned detection method and superpixel-based global contrast detection method, are pro-posed. These two methods are very good to complete the task in the primary stage.3) To select the most salient region fast and accurately, a new method for the advanced stage, key region focusing, is designed. The method can refine the basic saliency map of the primary stage by selecting some attended point and analyze their distribution. The key region focusing method can be combined with two methods of the primary stage respectively. Each new combined method, as a complete saliency detection method, can produce satisfying saliency maps for a variety of applications.4) Some saliency-based applications in image segmentation, image editing and other areas are implemented. For some proposed applications, new implementation methods are designed to achieve better effects. Some new applications are explored in the thesis, such as rapid targeting and background blurring. These new applications can make full use of existing technology and achieve impressive effects that is hard or impossible to do before.