Retrieval Of Cross-Modal Shoeprint Images Based On Depth Features

The research and application of machine learning in artificial intelligence has developed rapidly in recent years,which provides a lot of help for people’s life and work.The intelligent system combining image retrieval and shoeprint inspection provides a platform for criminal investigators to obtain relevant information about criminal suspects more accurately and quickly.For the problems of incomplete shoeprint and backward in retrieval technology,the deep learning is used to process the shoeprint images,which can accurately retrieve the extracted shoeprint images corresponding to the shoeprint from the discovery sites,and provide more clues and information for the investigation of the case.The use of this technology shares the burden of solving the case by judicial personnel and improves the efficiency of case resolution.In this thesis,a shoe-print image retrieval algorithm based on deep learning is studied.First of all,a cross-domain image retrieval method is proposed for the characteristics of the gap between the shoeprint images extracted from the crime scene and the shoeprint images in the gallery,and use the pre-process image segmentation of shoeprint images to reduce the inter-domain gap of images and optimize retrieval performance;secondly,in view of the feature that the environmental noise of shoeprint images has more influence,based on the image retrieval algorithm of deep features,we use the Siamese network as the basic network framework,the VGG19 network is used for the feature extraction,and the triple loss function is used to Similarity measures,to ensure that the extracted features contain more information about shoeprint and improve retrieval performance.Then an optimization scheme is proposed for the feature extraction and similarity measurement of the network.First,this thesis proposes a method to merge features extracted from two types of images,and then uses the merge feature to measure similarity.The features extracted from the original image contain all the shoeprint image information,while the features extracted from the segmented image has less noise impact,the merge feature effectively combines the advantages of the two,so the retrieval using the merge feature can further improve the retrieval performance.In this thesis,the triple loss is used in the similarity measure.The selection of positive and negative samples in the training set affects the performance of the network model.This paper proposes an online triplet mining method.By selecting negative example image that are more similar to the retrieved image,and selecting positive example image that have a greater gap from the retrieved image,in order to achieve better training results,and then obtain better retrieval results.