Research On Shoeprint Image Clustering Algorithm Based On Interactive Semantic Information Embedding

Shoeprint images are very important evidence,and they can be used to infer suspects and link similar cases.However,with the increase of cases,more and more shoeprint images are extracted from crime scenes,and those data are unlabeled.At present,content-based shoeprint automatic clustering methods often use the low-level feature representation,but the results always can’t meet the expectation of forensic experts for the semantic gap.The manual labeling method has high precision,but it costs labour and time.Therefore,we combine automatic labeling and a small amount of manual labeling to propose a shoeprint image clustering algorithm based on interactive semantic information embedding.The main work of this thesis is as follows.(1)An online semantic attribute prediction algorithm based on clustering guidance is proposed.Based on content-based shoeprint image clustering results,a small number of representative samples are automatically selected for interactive semantic attribute labeling.Through this part of the samples with semantic attributes online attribute learning,semantic attribute classifier is trained to predict the attributes of unlabeled samples.Experimental results on shoeprint images show the accuracy of overall attribute prediction is above 70%.(2)A shoeprint image clustering algorithm based on semantic attribute embedding is proposed.Semantic attribute is a bridge connecting high-level features and low-level features and is embedded in the whole process of shoeprint image clustering.The clustering results are redivided and recombined through the guidance of semantic attributes,so as to reduce the semantic gap and obtain clustering results that are more in line with what the user needs.Experimental results on shoeprint images show the F value reaches 80.37% and the purity reaches 98.50%.(3)A shoeprint image clustering algorithm based on multilevel graph representation and semantic attribute embedding is proposed.Traditional shoeprint image representation is often based on the content characteristics of a single image,ignoring the relationship between shoeprint pairs and the shoeprint semantic attributes.In this thesis,the characteristics of the image,the relationship between the images,and the semantic attribute of the image are combined to form a multi-level graph representation of the image.Then the multi-level graph representation of the shoeprint is introduced into the shoeprint image clustering algorithm based on semantic attribute embedding.Experimental results on shoeprint images show the F value reaches 83.07% and the purity reaches 98.76%.