Research On Translation Algorithm Based On Footprint Pressure Image

Compared with fingerprints and DNA information,footprints play an important role in the field of criminal detection and biomedicine because they are easier to obtain.With the development of science and technology,footprint recognition technology is also improving rapidly.However,most of footprint recognition algorithms are based on barefoot while the crime scene is mainly shoeprints.Shoeprints have different types and textures,which will bring great difficulties to recognition.How to mine barefoot footprints through shoeprints is one point of the key problems in the field of footprint recognition.Aiming at the above problems,the thesis constructs a hybrid dataset of barefoot pressure images and shoeprint pressure images,and then translate shoeprint pressure images into barefoot pressure images through deep learning technology,the main research contents are as follows:(1)Using deep learning technology to carry out Footprint translation experiments which requires a large amount of footprint data to support.Firstly,the plantar pressure data collection equipment is used to construct the first barefoot plantar pressure and sole pressure mixed dataset called SF2 DS,which supports footprint translation,footprint recognition and other tasks.Secondly,pre-processing algorithms such as denoising,segmentation,centering and rotation is proposed to improve the effect of subsequent translation experiments for the task of footprint translation.Finally,in order to verify the validity of the translation experiment,a method for evaluating the similarity of plantar pressure images is proposed to facilitate the subsequent evaluation of the translation experiment results.(2)In order to improve attention of Generative Adversarial Networks to the details of footprint images,a Symmetric Multi-branch Generative Adversarial Network(SMb GAN)is proposed.SMBGAN includes two feature pyramid generators,two local discriminators and a global discriminator,the generator is used to generate the target domain image,and the discriminator is used to judge the authenticity of the image.This strcture divides the footprint image into the upper part and the lower part for independent conversion.The translation experiments on SF2 DS prove that SMb GAN has the best effect in the footprint translation experiment.(3)Generative Adversarial Networks are prone to pattern collapse during the training process.In order to ensure the stability of the Generative Adversarial Networks model during the training process,the normalization of the spectrum is integrated into the structure of the Generative Adversarial Networks.At the same time,so as to enable the structure to autonomously learn the key areas between the footprints,the self-attention module is added in the generator and discriminator for enhancing the model’s attention to the key areas.Based on the above two structures,a generative adversarial network which combines Spectral normalization and Self-attention modules(Sa Sp GAN)is introduced.Experiments show that the training process of GAN with spectrum normalization is relatively stable,and the self-attention module could improve the effect of the footprint translation experiments.