Research On Key Point Location Method Of Palm Infrared Image

In traditional Chinese medicine,hand diagnosis is one of the characteristic diagnostic methods in traditional Chinese medicine inspection.It is an auxiliary diagnostic method that can infer the health status of human organs.It can judge the health status of human organs by the color,texture,shape and other characteristics of the palm Jiu gong area.This method is mainly based on the prior knowledge of traditional Chinese medicine practitioners.Firstly,the palm image is divided into nine palaces by positioning,and then each area is observed and diagnosed.Therefore,positioning error is easy to occur in the positioning process,which affects the doctor’s judgment of the disease.With the development of infrared thermal imaging technology,the thermal information of human palm can be obtained through the infrared thermal imager,and the thermal information plays an important role in the hand diagnosis of traditional Chinese medicine.The infrared thermal imager can make up for the deficiency of traditional Chinese medicine inspection,and analyze the corresponding viscera function state of human body and diagnose the health status.However,when using medical infrared thermograph to locate the key points of infrared image of the palm,because of the difference of the size and angle of the palm and the degree of stretching,the traditional positioning algorithm is difficult to achieve accurate positioning,which may lead to the problem of wrong diagnosis of doctors.Therefore,it is the first task to improve the positioning accuracy of the key points of the infrared image of the palm.In order to solve the above problems,this paper proposes a localization method of multiscale differentiated convolution feature pyramid fusion residual networks（MDFPF-Res Net）to locate the key points of palm infrared image.In order to make the model achieve better results,this paper uses the way of data enhancement to enlarge and reduce the data set,shift,flip,select and so on.Firstly,multi-scale hole convolution is integrated into the pyramid to improve the ability of image high-level semantic feature extraction.The hole rates are 1,6 and 12 respectively;Then,an improved Bottleneck module is added,and BN layer is used to improve the inaccurate positioning of multi-layer temperature profile;Finally,the multi-scale differentiated convolution feature pyramid（MDFP）level is fused with the output feature image of the improved bottleneck module,and the image features are extracted by repeated feature fusion,which improves the problem of missing key points in palm infrared image due to low spatial resolution.In order to verify the effectiveness of MDFFP-Res Net positioning,this paper conducts experimental comparison on the infrared palm image dataset through multiple networks.Experiments on accuracy,AP⁵⁰,AP⁷⁵ and AP^L show that the proposed method can accurately locate the infrared palm image,and the network has higher positioning accuracy compared with other networks.From the accuracy of the training process of the network,we can see that the network has strong robustness.The recall rate experiment shows that although the problem of missed detection may occur in the infrared palm image,the network is superior to the baseline model and other methods in processing the classification and recognition tasks of key points.The above experiment verifies the effectiveness of the proposed method in the location of key points in the infrared palm image.And through Tkinter library,a set of usable software is compiled.By using this software,the image can be positioned arbitrarily,and the positioning results are displayed in the interface.