Research On 2D Optical Barefoot Footprint Recognition Method Based On Similarity Analysis

Footprint is a kind of human biological characteristics,which is unique and repeatable.In the problem of footprint recognition,methods based on footprint theory need to rely on expert knowledge to extract features manually,and then analyze the similarity between footprints.This method has some limitations,and feature extraction is time-consuming and labor-consuming.Deep learning algorithm has strong ability of abstract feature extraction,and can be applied to footprint recognition.However,it is difficult to obtain a large number of footprint data from the same person,and the difference between categories is small,which brings challenges to the training of deep network.Aiming at the problem of 2D optical barefoot footprint recognition,this thesis constructs experimental datasets using 2D optical barefoot footprints.On the basis of few-shot learning algorithms and convolution neural network,two 2D optical barefoot footprint recognition algorithms based on similarity analysis are proposed and verified by experiments.The specific research contents are as follows:(1)2D optical barefoot footprint datasets is constructed.Relying on the footprint perception and analysis laboratory,according to the scientific and standardized collection process,the 2D optical barefoot footprint data of 139 people are collected,then datasets are constructed and preprocessed.In order to research into footprint recognition,this thesis adopts the 2D optical footprint acquisition instrument provided by Hangzhou Chancel Electronic Technology Co.,Ltd.for data collection.This thesis collects the effective 2D optical barefoot footprint data of subjects in normal walking state.At the same time,footprint images in the dataset are preprocessed to remove scale ruler and extract region of interest to remove noise and redundant information.(2)Based on relation network and multi-task learning method,a multi-task similarity network is proposed to realize 2D optical barefoot footprint recognition.Aiming at the problem of datasets only contain few images from the same person,an autoencoder is designed,and the multi-task similarity network is constructed by relation network and the autoencoder.Firstly,the encoder maps different footprint images to a common feature space;secondly,the feature maps of different footprints are concatenated;finally,utilizing the concatenated feature maps,the similarity module can calculate similarity scores between footprints.Due to the existence of autoencoder,the risk of over fitting in the network training process is greatly reduced,and the training effect of the network is improved.This thesis uses similarity scores to measure the similarity between support samples and test samples,so as to predict the label of the test samples,and achieves a good effect of footprint recognition.(3)Based on convolution neural network,a data pair similarity comparison network is proposed to recognize 2D optical barefoot footprints.In this thesis,convolution neural networks are used to extract abstract features with stronger expression ability,then the data pair similarity comparison network is designed.In this model,firstly,footprint images to be compared is used to construct footprint data pairs;secondly,the convolutional comparison features of the data pair is extracted by the comparison feature extraction module;finally,the similarity score computing module uses the convolutional comparison features to calculate the similarity scores between footprints.The experimental results show that the model can effectively identify footprints and achieve high accuracy.