Research On Shoeprint Retrieval Oriented Semantic Representation Method

Shoeprints are often found at crime scenes and provide valuable forensic evidence. Because of the existence of repeated offences, effective retrieval of crime scene shoeprints would enable investigators to not only link different crimes, but also identify potential suspects. Shoeprint management and retrieval have an important significance on solving cases.The semantic information of image usually correlates well with people’s subjective visual perception. The valid semantic information of image can improve the subjective and objective consistency of the retrieval results. In the case of low quality images, low-level feature based shoeprint retrieval methods can not always achieve high performance retrieval results. Shoeprint semantic information representation algorithm is studied. The main works of this thesis are as follows:(1) A shoeprint image semantic representation framework is proposed.The shoeprint semantic representation is classified into global based shoeprint semantic representation and local based shoeprint semantic representation. Field experts have done subjective tests to verify the reasonableness of the proposed framework.(2) A global shoeprint image semantic information extraction algorithm is proposed.The proposed algorithm includes three parts:1) The GVF model, prior knowledge and data fitting are used to extract the the whole contour of shoeprint. The proposed algorithm is robust to all kinds of shoeprint images with small damaged regions. The extracted shoeprint image contours have a good performance on subjective and objective consistency.2) On the basis of the experimental analysis on the characteristics of the shoeprint images contour, this thesis proposes two criteria to judge whether a shoeprint is left or right, the accuracy rate of judging the left or right foot by each kind of these principles alone is more than 96%. A bimodal Gaussian function with the confidence condition of those criteria in practical judgment results is added torefine the recognition result. The accuracy rate of judgment results is raised to 99%.3) On the basis of the analysis on correlative characteristics of shoeprint images texture direction, integral direction extraction algorithm of shoeprint image texture is proposed according to the distribution of the energy spectrum. The classification rate on the suspect database is more than 92%.(3) A local shoeprint image semantic information representation algorithm is proposed.The proposed algorithm includes two parts:1) A contour description algorithm which is based on sectional statistic information with log polar coordinate is proposed. The shoeprint semantic element is labeled using this extracted feature and SVM algorithm. The formalized representation and similarity measurement method of the semantic element are proposed. Better accuracy rate is obtained in different type of element marking experiments.2) According to the insufficiency of the element information representation ability, this thesis proposes two semantic unit representation styles and similarity measurement methods. The experimental results show that semantic unit representation has advantages in the process of retrieval.In order to test the performance of the shoeprint semantic information in the shoeprint image retrieval application,9294 shoeprint images of suspects are used as the gallery database. The recall rate of the top 10, top 20 and top 50 reach to 66.22%, 72.30% and 75.68% respectively. The recall rate of the retrieval results is further improved by combining traditional low-level feature based algorithm with the proposed semantic representation algorithm. The recall rate of the top 50 reaches to 78.38%. The experiment results show that the proposed shoeprint images semantic representation algorithm works well for improving the recall rate of the retrieval results and achieves better performance on keeping subjective and objective consistency.