Research On Interactive Image Segmentation Algorithm Based On Graph Cuts

Image segmentation is a fundamental and crucial work in the field of image analysis, pattern recognition and computer vision. Due to the complexity of natural image automatic segmentation methods are often poor generality and poor accuracy. By contrast, interactive or semi-automatic segmentation methods with intelligent optimization means can segment the target object quickly and accurately by using limited user interaction to get as much as possible segmentation information. So interactive methods have stronger practicability. Graph cuts is a kind of excellent intelligent optimization algorithm, has received the widespread attention in recent years. Interactive segmentation model based on graph cuts theory has been widely used due to combining region and boundary features perfectly and its excellent properties with multi-feature fusion, global optimum, higher efficiency, etc. It has become a new research hotpot in image segmentation field. In this paper, the model of interactive image segmentation based on graph cuts is optimized with three aspects by improving the efficiency of algorithm, multi-feature fusion and resolving shrinking bias problem. Our work mainly includes the following parts:(1)We propose a wavelet multi-scale iterative segmentation model based on graph cuts. GrabCut algorithm is an effective interactive image segmentation algorithm. But it uses iterative segmentation based on the whole image to estimate Gaussian Mixture Model (GMM) parameters, which seriously restricting the efficiency of the algorithm. To solve this problem, we propose a wavelet multi-scale iterative segmentation model based on graph cuts. We use the advantage of the multi-resolution analysis character of wavelet transform to optimize the GrabCut algorithm model. Through the wavelet transform of the image the low-frequency subband images are used as training samples to estimate GMM parameters with multi-scale iterative segmentation. By combining the easiness of coarse scale segmentation and accuracy of fine scale segmentation, we reduce the number of samples effectively and improve the efficiency of the algorithm on the premise of ensuring the accuracy. In addition, in order to solve the shrinking bias problem inherent in graph cuts we use high-frequency coefficients to detect multi-scale edges. The local adaptive regularization parameter is calculated with the edge probability map to improve the thin boundary image segmentation.(2)We propose a fast segmentation model of JPEG images based on graph cuts. In order to solve the problem of poor real-time performance for high resolution images of graph cuts algorithm we propose a fast segmentation model based on graph cuts for JPEG images which are widely used on the internet and by most digital camera manufacturers. We use the special coding format of JPEG image to extract the direct current (DC) coefficients to generate DC low-frequency image for estimating the GMM parameters, which reduce the number of training samples. Then the direct current and alternating current (AC) coefficients are used to generate texture features. Through calculating the KL distance between the distributions of color and texture features, these two features are effectively combined. The high frequency characteristic of AC coefficients is used to calculate the edge probability of pixels for constructing local adaptive regularization parameter. This model improves the segmentation efficiency of high resolution JPEG images and the ability for segmenting texture and thin boundary images due to the direct use of JPEG coding data without additional transformation.(3)We present an algorithm based on graph cuts using combined visual saliency features. Graph Cuts algorithm tends to produce segmentation errors and shrinking bias when the foreground and background color distributions overlap and uses edge probability of pixels to construct local adaptive regularization parameter which is difficult to determine boundary and needs additional calculation. To improve these problems we present an algorithm based on graph cuts using combined visual saliency features. With image saliency analysis and adding saliency constraint term in energy function the reliability of the data term constraint is enhanced. By constructing the mean saliency map and further processing for small areas the noise is reduced and the accuracy of saliency constraint is improved. Through estimating the degree of colors overlap adaptive adjustment of color term and saliency term is designed. Use the probability of pixel belonging to foreground/background directly to construct local adaptive regularization parameter which improves the efficiency of the algorithm. Our algorithm improves the segmentation result and shrinking bias phenomenon effectively.(4)We present a fast segmentation algorithm based on graph cuts using combined CS_LBP texture features. The interaction efficiency of graph cuts algorithm is not high due to the calculation based on pixel and when foreground and background colors are similar the values of data term are very close, so the value of energy function mainly depends on smooth term. Then the prior information obtained by user interaction will have limited usefulness which will result in errors of segmentation and when smooth term becomes more important the algorithm is prone to produce shrinking bias phenomenon due to the influence of the energy function minimization for thin boundary. So we must consider the important texture features. But there exists problem of big calculation and high complexity for co-occurrence matrix and Gabor filter which are often used to extract the texture features. Therefor, we propose a fast segmentation algorithm based on graph cuts using combined CS_LBP texture features. Mean Shift algorithm is applied to pre-segment the original image into regions to form superpixels and construct region adjacency graph which reduces the scale of s-t network flow graph effectively and improves the efficiency of the algorithm sufficiently. Then cumulative histogram, simple and effective CS_LBP texture descriptor are used to extract color and texture features from each region. A new term of texture constraint is added to the energy function and local adaptive regularization parameter is used. So the proposed algorithm can get better segmentation results with a significant efficiency improvement.