Attention Based Cross Pattern Shoeprint Sequence Identification

Shoeprint sequence refers to the collection of multiple shoeprints with temporal relationship,which can reflect the anatomical structure characteristics of the feet and biological information such as spatial geometric relationship and pressure characteristics during walking,and it has the characteristics of stability and not easy to be falsified.In banks,airports and other occasions where identity registration management are required,the shoeprint collection equipment can also be placed underground concealed and not easily detected by the subject.The use of shoeprint features for personnel identity verification and identification has high application value.The cross pattern shoeprint sequence refers to the shoeprint sequence images in the probe set and the shoeprint sequence images in the gallery set formed by the tread of shoes with different sole patterns.The difference in sole pattern can cause a large difference in the appearance of the same person’s shoeprint,resulting in increased difficulty in identity identification based on cross pattern shoeprint.For this reason,this thesis studies the cross pattern shoeprint sequence identification method,and the main work is as follows:1)A cross pattern shoeprint image preprocessing method is proposed.The incomplete shoeprint image is repaired by referring to the representative shoeprint image to solve the problem of shoeprint image defect caused by random interference.The pressure significance region of the shoeprint image is detected and extracted,and the fast marching method is used to predict the pressure value in the gap region of the shoeprint pattern in this region.Compared with directly repairing the entire shoeprint image,this method can reduce the errors in unimportant areas and the negative influence of the sole pattern.2)A cross pattern shoeprint sequence representation of the attention based mechanisms for shoeprint energy map set is proposed.Firstly,the shoeprint energy map sets are extracted for the pressure significance region to generate the shoeprint significance energy map sets,and then the energy entropy map sets and energy mean map sets are constructed by weighting and fusing the energy maps of different scales to generate a total of 4 types of energy map sets(24energy maps)to express the cross pattern shoeprint sequences.The recognition rates on dataset CSSD1,dataset CSSD2-T1,and dataset CSSD2-T2 are 79.35%,84.48%,and 92.81%,respectively.3)A cross pattern shoeprint sequence identification method based on convolutional neural network fusion channel attention is proposed.The method first uses the feature extraction module to get the feature maps of the shoeprint energy map set,then redistributes the feature weights of each feature map,and finally gets the identity prediction result by the identity prediction module.According to the characteristics of cross pattern shoeprint image,the data augmentation method of multi-scale truncated mean filtering is given.The recognition rates on dataset CSSD2-T1,and dataset CSSD2-T2 are 66.50% and 84.64%,respectively.4)A cross pattern shoeprint sequence identity identification method combined with stride features is proposed.Taking advantage of the simplicity of the stride feature extraction and the fact that only a single shoeprint sequence is required for registration and training,it is fused with other features to make up for the disadvantage that other features require multiple shoeprint sequences for registration,further improving the identification performance.The recognition rates on dataset CSSD1,dataset CSSD2-T1,and dataset CSSD2-T2 are 81.94%,86.21%,and96.08%,respectively.