Lip Region Segmentation Based On Markov Random Field

With the rapid development of computer science and technology,artificial intelligence has become a research hotspot in the information field.In a noisy environment,people's ability to perceive audio signals is reduced,so the way of voice interaction is limited.In recent years,the widespread application of computer vision has made a breakthrough beyond voice interaction.Researchers have found that using the variation of human lip movement characteristics can help solve speech content that cannot be recognized due to interference,namely lip reading technology.This paper mainly studies lip segmentation technology in lip language recognition system.Lip segmentation is a crucial preprocessing for computer vision and lip language recognition.It not only has important significance for the study of lip language recognition system,but also its segmentation result has a direct impact on the subsequent recognition rate.In terms of lip segmentation,this paper uses the Markov random field modeling method to classify pixels in the lip region.The Markov random field can sufficiently classify the target pixels by applying the spatial dependence of the pixels.Considering the gray scale of each pixel of the image as a random variable with a certain probability distribution,the segmentation problem is transformed into a label optimization problem by combining the maximum posterior probability criterion.For the wavelet-domain hidden Markov tree model algorithm,this algorithm provides a feasible solution for multi-scale statistical image modeling.Firstly,Haar wavelet decomposition is applied to the image,then the likelihood coefficients of the wavelet coefficients of each sub-band pixel are calculated,and the Markov tree structure is constructed by the node relationships in different frequency bands.The pixel-like likelihoods of wavelet coefficients with different resolutions are transformed into the correlations between different scales of wavelet coefficients to construct an optimization function.Finally,use the maximum expectation algorithm to estimate the function parameters.Due to the limitations of each scale segmentation of wavelet-domain Hidden Markov Tree Model,this paper adopts an improved multi-scale context information fusion method.This method reduces the influence of the inaccuracy of the boundary segmentation of each scale on the final result.The algorithm performs traversal from root node to child nodes and from child node to root node,captures the target free nodes and isolates the non-target classification area boundary nodes,so that the effect of interference factors such as lip boundary whiskers on the boundary connectivity.When the likelihood energy minimization of each scale is calculated,the classification information likelihood function and scale coefficients of the wavelet domain are included in the fine scale,making the segmentation result more accurate.Experiments show that the proposed method is feasible and can well segment the lip contour completely.Compared with the traditional conditional iteration model(ICM)proposed in this paper,the accuracy and smoothness of lip edge segmentation are improved.