The Research On Matting Algorithm Based On Prior Knowledge

Matting is the process of accurately extracting foreground objects from the background.Matting is a basic content of digital image processing and the basis of image composition.It is widely used in image editing,special effects production,and virtual reality.Complex details of object shape,complex background texture,alpha channel,disturbance from similar background and other factors make the matting extremely complicated.Since the image matting is a serious under-constrained problem,in order to obtain high-quality matting results,most matting algorithms require users to provide labeling information of the foreground and background.However,it is undoubtedly a complicated and time-consuming task for users to provide accurate foreground and background labeling information.How to ensure the robustness and accuracy of matting algorithm in complex environment still needs further study.This paper deeply analyzes the existing problems of image matting,focusing on the natural image matting algorithm and its problems under the guidance of prior information.This paper starts from the optimization of user input,trying to study more convenient interaction ways,reducing user interaction workload while providing more accurate foreground and background labeling information,in order to obtain high-quality matting results.The accurate trimap is the guarantee of the performance of the matting algorithm.In order to solve the problem that the unknown region of the trimap is too large,the priori information provided is not accurate and sufficient,and the human interaction is complex and time-consuming,this paper studies how to create accurate trimap quickly.The main research contents of this paper include:(1)This paper proposed the creation of trimap based on popular manifold ranking algorithm.The manifold ranking algorithm relies heavily on the query nodes of labels.In this paper,the foreground/background query nodes are obtained by simple manual labels.In order to obtain a better computing environment,this paper smoothes the input image to enhance the foreground objects while removing the interference of unnecessary details and textures.This paper sorts foreground superpixels and background superpixels separately.Manifold ranking algorithms will give each superpixel a different sorting score.In this paper,we calculate the threshold of superpixels and combine the results of foreground ordering and background ordering to create an accurate three-part map.The experimental results show that the method in this paper can generate high quality trimaps,thus ensuring the accuracy of the alpha masks that are estimated by the matting algorithm.(2)This paper proposed a matting algorithms based on gradient sparseness.Complex background and blurred foreground/background boundary will affect matting algorithm performance.Complex textures make it difficult to accurately extract foreground objects.This paper use gradient sparse priors to divide the image into two layers and normalize the gradient of the two layers so that one layer has a long-tail distribution and the other has a short-tail distribution.The gradient of the two-layer image does not have consistency.This paper gets a clear foreground/background boundary by suppressing complex background textures and highlighting foreground objects.Experiments have shown that we can get a higher-accuracy alpha mask by processing the image with gradient sparse priors.(3)This paper proposed a method for assessing the robustness of matting algorithms based on Gaussian-Hermite moments.Most matting algorithm evaluation methods mainly focus on the quality of alpha mask but do not pay attention to the robustness of the algorithm,and the evaluation results may be inconsistent with people's subjective feelings.The loworder moment of the image can better capture the image features,the gradient direction can describe the representation and shape of local objects in the image,and the gradient amplitude can accurately reflect the contrast and texture changes of tiny details in the image.This paper chose Gaussian blur,salt-pepper noise,and combination blur-noise to destroy the image,and then evaluated the consistency of the original alpha mask and the noise alpha mask.The experimental results show that the noise will have a great impact on the performance of the matting algorithm.The performance of matting algorithms shows a downward trend as the noise level in the image is enhanced.Different matting algorithms have different degrees of adaptation to different noises.In the robustness evaluation,the performance of traditional matting algorithms outperforms even recently proposed advanced matting algorithms.