Shoe Print Retrieval Algorithm Based On Deep Learning

Shoe prints are the most common trace evidence at crime scenes.Investigators can analyze information such as the size and type of shoes worn by the perpetrator based on their footprints left at the crime scene.This not only provides a basis for the combination of cases,but also becomes an important part of the "surveillance+shoe printing" technology in recent years to identify suspects.Based on the image retrieval and comparison of on-site shoe prints,the shoe type is determined as shoe print retrieval.Currently,shoe print retrieval still faces difficulties such as low image quality and low algorithm accuracy in the field.In response to this issue,the following work has been carried out in this article:Firstly,build a shoe print retrieval database.This article establishes a shoe print dataset SP Database that is more suitable for training and testing shoe print retrieval algorithms.The dataset includes 608 categories of original shoe print images,totaling 2302 pieces.Among them,there are 408 types of training sets,totaling 1702 sheets;The test set consists of 200 categories and a total of 600 sheets.Secondly,propose a shoe print retrieval algorithm based on an improved lightweight network Efficient Net.This article selects Efficient Net-B3,a lightweight network with excellent classification performance and low computational cost,as the feature extraction network.By replacing the Coordinate Attention attention mechanism,using the Fused MBConv module in the shallow layer of the network,and using the Dilated MBConv module in the deep layer of the network,the training speed and retrieval effect of the algorithm are improved.Thirdly,propose a shoe print retrieval algorithm based on a multi-scale global shoe print feature extraction network.In terms of global shoe print feature extraction,the algorithm was improved using multi-scale feature fusion,and the information contained in the fused features was further optimized using a fast normalization weighted fusion method.In the training phase,the loss function is optimized,that is,the label smoothing cross entropy loss assisted hard sample sampling triple loss function.Fourthly,propose a shoe print retrieval algorithm based on a multi part shoe print feature extraction network.On the local footprint feature extraction,the PCB local footprint feature extraction module is added,and the multi branch loss function joint expression is used as the loss function of the entire algorithm in the training phase;In the testing phase,the extracted local shoe print features are concatenated with the global shoe print features as descriptors for shoe print retrieval,further improving the retrieval effect of incomplete shoe prints.Through the above work,this article ultimately achieved retrieval accuracy of 95.5% and100% for top 1% and top 10%,92.5% and 98%,77.73% and 91.28%,95.45% and 98.48%,respectively,on the four datasets of CSS-200,SP Database,FID-300,and CS Database(Dust).This indicates that the algorithm in this article effectively improves the retrieval accuracy of incomplete fuzzy shoe prints,and is more suitable for criminal investigation in public security operations.