Lip Segmentation Based On Deep Fuzzy Convolutional Network

In recent years,with the rapid development of pattern recognition technology and speed of computer operation,more and more deep learning technologies have been applied to various fields.Deep learning-based lip reading technology has also received attention from scholars domestic and abroad for its excellent recognition performance.The purpose of lip reading is to use the visual information of the lip image sequence to supplement the auditory information such as speech,thereby improving the computer system's ability to understand language.Lip image segmentation facilitates the automatic extraction of information such as the contours of lips and directly affects the performance of subsequent speech recognition modules.Therefore,achieving accurate pixel-level lip region image segmentation is of great value to the application of lip reading technology.Most existing lip segmentation methods can generally achieve better results in a laboratory environment with a fixed shooting scene.However,under natural scenes with complex lighting and backgrounds,when the speaker opens his mouth to expose the mouth,tongue,gums,and other oral components that are very close to the color of the lips,the performance of various segmentation algorithms will be significantly reduced for such captured images.At the same time,traditional algorithms have high computational complexity and cannot meet the real-time requirements of mobile applications.In view of the above problems,this paper proposes a lip image segmentation technology based on fuzzy deep convolutional neural network,which seamlessly fuses the deep convolutional network and fuzzy logic modules to achieve accurate,robust and efficient pixels in various difficult scenarios.Convolutional neural networks can provide multi-level image feature expressions and provide sufficient local and global information for lip segmentation;while fuzzy logic modules can fully model the uncertainties generated by lip images during shooting,labeling,and other links,improving This improves the robustness of the segmentation algorithm.In addition,this paper also designs a knowledge distillation technology for lip image segmentation,which makes full use of the supervised information of the large model to improve the segmentation accuracy of the small model,and then completes the model compression to improve the real-time performance of the segmentation algorithm.Finally,this paper constructs a lip image dataset under natural scenes.After experimental demonstration,compared with existing lip image segmentation algorithms,the proposed method achieves higher accuracy while satisfying real-time performance.