Adaptive Appearance Separation For Interactive Image Segmentation

Nowadays,our society is in the era of data explosion,among which,the number of images is amazing.In order to analyze the image data and gain further understanding,image segmentation is an indispensable step.Because of its importance,a large number of new image segmentation al-gorithms are emerging.The purpose of image segmentation is to segment the objects(i.e.,targets)associated with the real world in the image.How-ever,because the image resolution becomes larger and the image contains more and more rich contents,people put forward higher requirements on the performance and effectiveness of the image segmentation algorithm.Many images in the real world can be complicated.What the objects in the image of the user perceive cannot often be reflected by the result of a fully automated image segmentation algorithm.Therefore,by user inter-vention-specifying some pixels belong to the target,some pixels belong to the background,the interactive image segmentation incorporated with the understanding of the image content of high-level is a better choice.Al-though there are many semantic segmentation algorithms based on deep learning,most of these methods require a large number of segmentation sample results with specific size and high-performance hardware resources as a prerequisite for their learning,in addition,they also need to elaborate-ly design the neural network architecture and make further efforts to train and tune the model to achieve the segmentation of accuracy that is not too high.This paper focuses on the interactive image segmentation algorithm that does not depend on a particular dataset.It can accurately and effi-ciently segment the target according to the subjective perception of human beings without a requirement for the pre-segmented ground truth image in advance(only as a reference to evaluate the segmentation accuracy).The segmentation results obtained by the algorithms in this paper can be di-rectly used in medical diagnosis,objects recognition or tracking,etc.,and can also be directly used as the labeled sample for semantic segmentation of deep learning.This paper puts forward the appearance separation model for inter-active image segmentation,and studies this model that applied to dense condition field(Dense CRF)framework and graph-cut framework used for image segmentation respectively.The algorithm finds and solves the problems of the two traditional algorithms,and does not need the iterated optimization process multiple times,so as to realize an efficient and accu-rate foreground-background segmentation.The main innovations of this paper are as follows:1.A novel appearance separation model for interactive image segmen-tation is raised.The model can reduce or even eliminate the clutter from the background and get a proper initial soft segmentation by combining the difference of the foreground and the background color feature and the corresponding difference of the geodesic distance in-formation.It takes a big step for the realization of the finally accurate segmentation.In the real world,there are many foreground-back-ground clutter or appearance similar images;common image seg-mentation algorithm can hardly get a satisfactory segmentation result due to the interference of light,texture,and blurring.The appear-ance separation model proposed in this paper makes full use of the color information and edge information,adds the human-level image understanding through spatial location information provided by the user,and achieves the foreground-background appearance separation so as to make the complex image more accurate segmentation.2.A precise image segmentation algorithm based on Dense CRF frame-work with appearance separation model is proposed.Classical algo-rithms based on the Dense CRF framework,such as the DenseCut algorithm,has as many as seven human-specified parameters that re-quire either complex parameter learning or time-consuming statis-tics,all depending on the particular dataset.In addition,they may require multiple inferences of the mean-field for iterative optimiza-tion to achieve satisfactory segmentation results.Moreover,Dense-Cut for seeded interactive mode often requires a large amount of us-er interaction workload for images that are apparently cluttered or have similar foreground and background color features.According to the characteristics of the image to be segmented,the algorithm in this paper adaptively sets and fine-tunes the related parameters in the Dense CRF framework.Applying the proposed appearance separa-tion model to the framework,a more accurate segmentation results can be obtained after a mean-field inference.The algorithm obtained in this paper reduces the user scribble interaction workload.Exper-iments on three famous datasets show that the improved DenseCut algorithm after adding the appearance separation model proposed in this paper is superior to the five interactive segmentation algorithms proposed in recent years on the segmentation accuracy,and is close to or less than them regarding segmentation time.3.A fast image segmentation algorithm combining appearance separa-tion,appearance overlapping penalty and Graph-Cut framework is put forward.The OneCut algorithm based on appearance overlap avoids the drawbacks that the classical GrabCut algorithm needs mul-tiple iterations of graph cuts and greatly improves segmentation ef-ficiency.However,after running plenty of experiments,I found that OneCut algorithm is prone to a large number of isolated points,and the number of isolated points is positively related to the weight of the appearance overlap penalty item under certain conditions.The appearance separation model proposed in this paper and appearance overlapping information are reasonably integrated into the efficien-t graph-cut framework,and a reasonable overlap penalty weighting scheme is proposed to alleviate the OneCut algorithm tends to ap-pear isolated points.Besides,the improved OneCut improves its seg-mentation accuracy and reduces the user's interactive workload.The segmentation experiments on three internationally accepted datasets demonstrate the improvement of the algorithm proposed in this paper.After comparing the improved OneCut algorithm with three other re-cent interactive image segmentation algorithms on the efficiency and accuracy,I find that my method achieves the highest precision,and the algorithm is efficient and feasible.