Studies On Image Segmentation Method Based On Saliency

Image segmentation is one of the hotspots in the field of computer vision,and it is the critical stage in image processing.Image segmentation is based on the similarity principle of pixel features in the image,and the image is divided into several disjoint regions,so that the features of the image pixels in the same region are similar,and the similarity of pixels between different regions is very low.The results of image segmentation affect the scene understanding and target recognition in the subsequent stage of image processing.With the continuous development of modern science and technology,people get more and more means to obtain images.More image segmentation algorithms have been proposed to be applied to mobile devices or the internet.Although thousands of algorithms have been proposed since image segmentation,However most of these algorithms have some problems.Most of the traditional automatic segmentation algorithms do not take into account the special visual processing of the human eye in the observation of the image,so the image segmentation results are sometimes very different from the expected results,the interactive image segmentation methods require manual initialization,in the image of the foreground and background color similarity,The interactive algorithm is difficult to give a reasonable result of image segmentation.To solve these problems,through the collection of relevant information at home and abroad,based on in-depth study of image segmentation algorithm and visual saliency algorithm,effectively solves the problem that the initialization area of the Grabcut algorithm can not be automatically extracted.,foreground and background colors are similar to the error segmentation problem.The algorithm is improved slightly and applied to the problem of flame target detection.The main research work of this paper is as follows:(1)In order to overcome disadvantages of the Grabcut algorithm in artificial initialization sensitive region problems and in the complex background,the Grabcut algorithm is easy to cause the problem of error segmentation.This paper studies how to use visual saliency map to automatically extract the initialization area and effectively reduce the error segmentation when the foreground and background color of the image are similar.TA novel Grabcut algorithm is proposed by combining the saliency detection method and the Grabcut algorithm,This method first obtains a better image salience by improving the saliency measure;then the saliency map is converted into a constraint to improve the reliability of regional items added to the Grabcut area in the end;instead of manual interaction using the saliency map,the initialization of Grabcut,Finally,to realize the automatic segmentation of Grabcut and reduce the segmentation error of Grabcut algorithm when the foreground and background color are similar.In this paper,using the MSRA1000 data set for experiment to compare and analysis,The experiment is tested from two aspects of subjective vision and objective performance.Compared with the IT algorithm(Saliency detection based on biological model),GBVS algorithm(Graph Based Visual Saliency),SR algorithmI(Spectral Residual),FT algorithm(Frequency-tuned Salient Region Detection)and Grabcut algorithm,the 5 algorithms are compared and analyzed.The experimental results show that this method can be better than the other algorithms from the visual effect.In terms of objective performance,the accuracy and recall of this algorithm on the MSRA dataset are higher than those of several other algorithms.(2)The improved algorithm is applied to the detection of flame target in single frame image: when the flame detection,sometimes within a short time the flame will instantly become larger,which makes the video changes before and after a few frames of intense,Continuous frame processing is difficult to detect flame targets accurately.Flame detection for single frame images using visual saliency method.Flame detection is carried out from two aspects of frequency domain and airspace.The two salient images are fused to form the approximate flame target region,and then Grabcut algorithm is used to accurately segment the flame target.The two ready-made data sets available on the Internet are FOTOSEARCH and WILDFIRE,respectively.In ordinary flame scenes,smoke flame scenes,multi-heaps of relatively compact flame scenes,and Hidden Flame scenes,the experimental results show that the algorithm in this paper can better obtain the significant target in the single frame image.