Research On Eye Acupoint Location Algorithm Based On Generative Adversarial Networks

With the wide spread and application of traditional Chinese medicine in the world,acupuncture technology plays an increasingly important role in disease treatment and physical health care.Quick and accurate acupoint positioning is a prerequisite for the implementation of acupuncture.Traditional acupoint positioning methods need sufficient practical experience to achieve accurate and fast acupoint positioning.In order to get a wider range of application of acupuncture technology,the automatic location of acupoints has become an important core technology.Due to the limitation of the complexity of acupoint location and the development of technical means,the existing automatic location methods of acupoints have some drawbacks,such as large positioning error,weak generalization ability of algorithms,and complex operation,which cannot meet the needs of large-scale clinical application of Chinese medicine acupuncture.Because the current research on the orientation of acupoints is small and there is not a large amount of data for reference,it is a challenging topic to automatically locate the important acupoint positions in the eyes.In recent years,with the progress of science and technology,the expansion of datasets and the advancement of image processing unit(GPU),machine vision,especially in-depth learning,has entered a leap-forward development.Image generation and unsupervised training mechanism based on Generative Adversarial Networks(GAN)network model have become a research hotspot in recent years.This paper mainly studies the method of eye acupoint positioning based on the generation of antagonistic network,introduces the idea of generating antagonism into the automatic positioning of eye acupoints,and turns the problem of acupoint positioning into the problem of generating image with acupoint information.Network training is conducted through interactive antagonism between generator and discriminator.The main research work is as follows:1.Considering that GAN network requires a large amount of data for model training,it is found that there is no facial acupoint dataset available for in-depth learning network training at home and abroad.Therefore,300 volunteers were recruited to collect facial images and a set of matched eye acupoint image datasets was constructed for network training.Experiments have proved that the dataset can achieve good acupoint localization effect when used in network training.2.In view of the simple structure,unstable training and difficulty in achieving a balanced state of the original Pix2pix network,an acupoint graph estimation network is proposed to locate the ocular acupoints.This design can greatly reduce the size of the update parameters required by the generator.Compared with the traditional Pix2pix network,the proposed acupoint map estimation network has better performance than the traditional Pix2pix network in point positioning,and the positioning training speed has certain advantages.3.An eye acupoint location network based on improved CycleGAN network is presented.Compared with the single image generation mode of Pix2pix,the double-loop generation mode of CycleGAN network can make full use of the antagonistic advantage of generator and discriminator,and subtly restrict the image by the loss of cycle consistency.The generator can better extract the characteristics of acupoint images by the inter-domain conversion of the image.The experimental results show that the ocular acupoint location network based on the improved CycleGAN network achieves better positioning performance,but the positioning speed is lower than the acupoint graph estimation network proposed above.4.To solve the problem of insufficient image data of eye acupoints,a network of acupoint information separation and fusion using unsupervised training method is proposed.The acupoint feature information is extracted from the image with acupoint information.The decoder end of the generator and the clean face feature image are fused at multiple scales.The acupoint information is integrated into the face image,and the automatic location of acupoints under the condition of insufficient training data is achieved.5.In order to evaluate the results of acupoint positioning,this paper defines two evaluation indexes of acupoint positioning performance,average coordinate error and point-out rate,so that the matching number and location accuracy of acupoint positioning can be objectively evaluated.