Research On Single Image-Based 3D Ear Reconstruction And Auricular Point Localization

The traditional Chinese medicine ear acupoint therapy is a treatment method that uses the reflex zones in the ear to treat various diseases,which has significant therapeutic effects such as adjusting human physiological functions,relieving pain,and improving conditions.Due to the small area of the auricle and the large number of acupoints in the ear,the accurate positioning of ear acupoints requires high experience for physicians,and the learning difficulty is also high,which may lead to limited inheritance of ear acupoints.This article proposes to use 3D reconstruction based on a single ear image and ear acupoint subdivision positioning to assist in the development of ear acupoint therapy,including a standardized collection of ear images,personalized reconstruction of ear models,and precise localization of ear acupoint key points.The main work of this article is as follows:(1)Clear internal ear structure images are the prerequisite for ear acupoint segmentation positioning.Given the unclear internal ear structure in the currently available public datasets and the lack of any related labeling points,this article proposes a standardized process for ear image acquisition.We obtained ear images through selfcollection and labeled ear acupoint-related characteristic points under the guidance of professional Chinese medicine physicians,ultimately forming a traditional Chinese medicine ear acupoint dataset.(2)To reduce the interference of background images as much as possible,this article proposes a method based on an improved YOLOv5 s network to detect and locate the ear contour region in the image.First,we used a Gaussian mixture model to coarsely segment the ear skin and facial skin and added a coordinate attention mechanism in YOLOv5 s to better focus on the global information of the ear contour during the convolution process.Finally,we used the YOLOv5s-CA network to train the model to achieve accurate detection of the ear contour area.(3)Considering the directional differences between left and right ears in model training,this article proposes an ear acupoint localization model with an added ear normalization module.By normalizing the ear contour direction in the ear image to the same side before model training,we solved the problem of ear direction differences.Combined with the Mobile Net V2 model,we achieved accurate positioning of 91 ear acupoint characteristic points in 2D ear images.Experiments show that this method can also achieve real-time positioning of the ear contour in video streams.(4)Considering the limitations of ear acupoint localization on 2D images in practical scenarios,this article proposes a 3D ear model reconstruction method based on single ear images,which enables the viewing of ear acupoint characteristic points from multiple angles.Based on a deformable human ear model and using the Res Net50 convolutional neural network as the main network,we added a channel attention mechanism in the model and used a weight loss function for feature points to establish a mapping relationship between 2D images and 3D models,thus achieving 3D ear model reconstruction and ear acupoint localization and segmentation.