Research On Image Foreground Region Extraction Algorithm Based On Geodesic Distance

Image foreground extraction is a hot and challenging research topic in the field of computer vision,and it is also a preprocessing technique for many image research fields such as search classification,target recognition,and image segmentation.The goal of image foreground extraction is to divide an image into foreground and background regions.It can acquire the parts of interest and the required features in the image,and precisely define the boundaries of the region of interest.It has been widely applied to various aspects of image research fields.However,at present,most foreground extraction methods are sensitive to extreme conditions such as noise and lighting,and it is difficult to accurately extract the segmentation boundary and ensure the integrity of the object.Therefore,this research takes the natural image of complex background as the research object,takes the key problems and difficult problems in image foreground extraction as the starting point,and combines the basic features of image salient targets,we conducted in-depth research and proposed two new methods of prospect extraction from the two aspects of automatic extraction and interactive extraction of the target.At the same time,in order to extract the salient targets in the image,we also propose a saliency detection method based on probability estimation,which is used in the extraction of interactive objects.The main research works are as follows:1.For the lack of automatic seed selection in the traditional region growth when the foreground is extracted,an automatic seed selection method combined with scan lines is proposed.At the same time,automatic extraction based on significance detection is completed.The main contents are as follows: Under the precondition of using superpixel segmentation algorithm to preprocess the image,we use the SGC saliency detection method to locate the significant target object,determine its centroid and background color set,and then use the scan line to automatically select seeds from the background color set.Then combine the scan lines to automatically select the seeds from the background color set for region growing.Finally,we achieve the object extraction by using the open operation and the isolated area disambiguation algorithm.The experimental results show that the algorithm is simple and fast,and the outline of the extracted object area basically fits the boundary.2.Aiming at the boundary ambiguity in the salient regions obtained by most saliency detection algorithms,a saliency detection method based on geodesic distance and probability estimation was proposed.The method firstly uses the superpixel segmentation algorithm to preprocess the image;secondly,it uses the “boundary prior” knowledge: that is,most of the superpixels at the image boundary belong to the background.Based on this,we construct a background probability estimation model and come to the probability that each superpixel belongs to the background;finally,the Laplace smoothing method is used to optimizes the saliency map,ensuring the consistency within salient region.The experimental results show that the saliency region detected by the method in our work is more uniform and the boundaries are more clear;at the same time,the background area is well suppressed.3.Currently,there are a lot of background interactions in the interactive image foreground extraction algorithm.In addition,when the saliency object is not obvious or the colors near the foreground boundary are nearly similar,the extraction result will be worse.To solve the above problems,a method based on geodesic distance is proposed.Firstly,on the basis of superpixel segmentation,only some strokes are alternately drawn on the target to be extracted,and the selected superpixel is taken as a known partial foreground target,and applied to the probability estimation method proposed in Research 2 to estimate the probability that the superpixels belong to the foreground.Then we use the obtained probability to construct the Graph Cut energy function and complete the initial segmentation of the target.Finally,this paper proposes a dynamic gradient diffusion method.This method can adaptively diffuse a gradient according to the gradient at the target boundary.The border zone is used to correct the results of Graph Cut segmentation.All the experiments in our work were performed on the MSRA10 k and ASD datasets,and compared with other different methods to verify the validity of the results.