Local Semantic Filter Bank Based Low Quality Shoeprint Image Retrieval

The shoeprint image is extracted from the crime scene and is one of the most common and important evidences in the crime scene,so it is important to the investigators.The purpose of shoeprint retrieval is to retrieve the shoeprint image that is most similar to the shoeprint image at the crime scene from the shoeprint database,so as to help criminal investigators to reveal clues about the criminal case.Since the shoeprints collected from the crime scene are usually defective and blurred low-quality shoeprint images,it is a great challenge for shoeprint retrieval.The existing algorithms of shoeprint image retrieval still have low robustness to low-quality shoeprint images.Therefore,this thesis proposes a low-quality shoeprint retrieval algorithm based on the local semantic filter bank.The main works are as follows:1)Local semantic filter bank construction algorithm is proposed.In this method,this thesis presents an interactive extraction rule for different types of local semantic patches.By accurately extracting the periodic and marked local semantic information existing in low-quality shoeprint images,the semantic gap problem is effectively improved.Then,the method establishes different local semantic filter models for different local semantic information and constructs a local semantic filter bank,thereby reducing the noise interference in the shoeprint images.Therefore,this method is more robust to low-quality shoeprint images,and the retrieval accuracy is improved.After that,this thesis fuses the similarity scores based on the local semantic filter bank and the Wavelet-Fourier Mellin feature to get the final ranking scores.In this thesis,the evaluation experiments are carried out on three shoeprint databases,namely MUES-SR10KS2S,FID-300 and CS,all of which have achieved good retrieval performances.The cumulative matching score of the top 2%on the low-quality crime scene shoeprint database MUES-SR10KS2S is 92.5%.2)A local semantic similarity enhancement algorithm based on pattern type is proposed.According to the pattern type of the query image,the transformation function is designed to enhance the local similarity,thereby effectively improving the retrieval effect of low-quality shoeprint images.In this thesis,the evaluation experiments are carried out on three shoeprint databases,namely MUES-SR10KS2S,FID-300 and CS,all of which have achieved good retrieval performances.The cumulative matching score of the top 2%on the low-quality crime scene shoeprint database MUES-SR10KS2S is 93.5%.3)A multi-similarity adaptive fusion strategy based on the area of the closed region of similarity curve is proposed.This thesis adaptively adjusts the similarity weight coefficients between the two database images based on the enclosed area of the similarity curve with different features and introduces the structure of the manifold sorting algorithm to improve the robustness of the shoeprint retrieval algorithm.In this thesis,the evaluation experiments are carried out on three shoeprint databases,namely MUES-SR10KS2S,FID-300 and CS,all of which have achieved good retrieval performances.The cumulative matching score of the top 2%on the low-quality crime scene shoeprint database MUES-SR10KS2S is 95.2%.