Research On Tongue Feature Recognition Based On Image Segmentation And Detection

These years,deep learning techniques have been used in Chinese medicine tongue diagnosis widely,mainly applied to semantic segmentation and feature analysis of tongue images.However,due to the complexity of face images,tongue segmentation is susceptible to lip and facial inhomogeneities,leading to some difficulties in tongue body segmentation.Due to the diversity of its morphology and location,some tooth marks and cracks may be very small,and the recognition of tongue tooth marks and cracks is more difficult.In addition,tongue data acquisition and annotation are difficult,and the annotation process requires some background experience in TCM.Therefore,solving these difficult problems is still one of the challenges to be faced in the application of deep learning techniques in the field of TCM tongue diagnosis.The purpose of this thesis is to investigate how to use deep learning techniques to solve the above problems,and this thesis will focus on how to improve the effectiveness of tongue segmentation and feature extraction.The main contributions of this thesis include:1.A home-made dataset of tongue diagnosis semantic segmentation and cracked tooth marks features was produced.4500 tongue semantic segmentation data of different age groups and health states,as well as 3206 data with tooth marks and 3742 data with cracks were obtained from the original images provided by Chinese medicine hospitals,after screening,pre-processing and annotation under the guidance of Chinese medicine practitioners and other operations.2.A segmentation algorithm based on squeeze excitation jump connection and multiscale fusion is proposed for tongue segmentation in Chinese medicine tongue diagnosis.The structure uses Res Net50 as the encoder with fusion attention mechanism in order to solve the feature loss problem of the encoder due to pooling.In order to solve the problem of poor fusion effect with large feature differences of encoder decoder,residual jump connection and deep squeezed convolution module are proposed in this thesis.Meanwhile,the multi-scale feature fusion module is used to combine the feature information to further improve the segmentation accuracy.The experimental results show that the algorithm achieves 94.56% m AP in the self-made tongue diagnosis segmentation dataset,which is2.02% higher than the U-Net algorithm.3.An attention-based cross-layer feature fusion target detection algorithm is proposed for the recognition of tooth marks and cracks features in Chinese medicine tongue diagnosis.This thesis proposes the feature fusion module CBAC,which introduces an attention mechanism and uses jump connections to better fuse backbone and neck features.To capture defect regions more accurately and filter out useless information,this thesis also proposes the attention NECA module,which is placed in the backbone network to enhance feature extraction and connect different size output branches of the neck to establish multi-scale feature map connections.Considering the dense distribution of tongue and teeth marks and cracks,which may cause the generated anchor frames to obscure each other,the DIo U loss function is proposed in this thesis instead of the original GIo U loss function to further improve the detection effect.The experimental results show that this algorithm is applicable to the self-made tongue diagnosis detection dataset m AP@0.5Reached 89.19%,an improvement of 2.30% compared to the YOLOv5 algorithm.4.Design and implement a traditional Chinese medicine self-service tongue diagnosis system,using the improved algorithm proposed in this article to achieve some functions.Recognize feature information in tongue images,provide relevant diagnostic suggestions,and enable users to diagnose independently through the system.