Research On Salient Region Detection Based On Superpixel

With the rapid development of the information acquisition technology and multimedia processing technology, Human beings have entered the information age. Massive digital media brings many conveniences to people’s daily life, but also puts forward a new challenge for computer vision processing technology. Because the computer’s ability to deal with the huge amounts of information is very limited, human beings urgently want that the computers, like human, has ability to choose a small amount of important information from a large number of visual information and to analyze and process in detail. Visual salient region detection technology can automatically predict, position and dig for the most important visual information in a scene,and can be also effective for visual media information filtering with the help of computer. There are many footholds in image salient detection methods, such as pixels, regional, superpixels, and so on. At present, the focus of salient detection is based on the super pixels, which is also a development trend in the future. In view of this, this thesis uses super pixels combined with other salient detection methods to acquire the final salient figure. Main work includes the following two parts:Firstly, a saliency region detection algorithm based on S LIC superpixel and Bias framework is proposed. In order to reduce the computation complexity, the SLIC algorithm is used to extract the super pixels of a given image in the image preprocessing. Next, at each scale, three essential cues such as local contrast, integrity and center bias are considered within the Bayesian framework. Then, the saliency figure is computed by weighted summation and normalization. Finally, the final saliency figure is optimized by a guided filter in order to further improve the detection results. A series of relevant qualitative and quantitative experiments are done on a MSRA10 K dataset containing 1,000 test images with ground truths, and results show that the performance of the proposed saliency algorithm is better than the current popular methods.Secondly, in view of the low precision problem of the traditional saliency region detection approaches based on image low-level features, a saliency region detection algorithm based on SLIC super pixel and global contrast is proposed. In order to reduce the computation complexity, the SLIC algorithm is used to extract the super pixels of a given image in the image preprocessing. Then the characteristics of the image global contrast and LBP texture feature are taken into account in the CIELAB color space, and the final saliency map is acquired by combining the two features with the MKB algorithm. Finally, the final saliency figure is optimized by a guided filter in order to further improve the detection results. A series of qualitative and quantitative experiments are done on a MSRA10 K dataset with ground truths, and simulation results show that the performance of the proposed saliency algorithm is better than the prevailing current methods.