Local Semantic Patch And Manifold Ranking Based Shoeprint Retrieval

A shoeprint is one of the evidences of high crime scene retention rate.Due to the repeated crimes by the same person,shoeprints left on the scene can help investigators to quickly link several cases,which can greatly increase the probability of detection.At the same time,if there are shoeprints of the related person in the criminal investigation database,the scope of the case can be directly limited,and even the suspect can be directly locked.Therefore,the efficiency and effectiveness of crime scene shoeprint retrieval is a technical problem that needs to be solved urgently in the field of criminal investigation.Most existing shoeprint retrieval algorithms don't take into account local regions which users are interested in.They extract features from the entire shoeprints firstly,and then perform retrieval based on the extracted features.Because most of crime scene shoeprints are incomplete and blurred,some local regions should be paid more attention.Moreover,the existing shoeprints reranking retrieval algorithm is inaccurate for the relationship between the partial dataset images,so it needs to be improved.This paper proposes a shoeprint retrieval algorithm based on local semantic patches and improved manifold ranking.The main works are as follows:1)A Local semantic patch based crime scene shoeprint retrieval algorithm is proposedThe crime scene shoeprints are usually incomplete and blurred.In order to alleviate the impact of these phenomena on retrieval performance,this thesis not only uses the entire image as input,but also introduces a local semantic patch.This thesis proposes a local semantic patch based shoeprint retrieval algorithm,and the algorithm gives two rules:preferentially selecting clear local regions and preferentially selecting repetitive local regions.According to these two rules,the user is guided to select a local region on a query image,and the region is called local semantic patch.Then,the local semantic patch and Wavelet-Fourier Mellin features are merged at the score level,and the fused score is used as the ranking score.Experiments were conducted on the crime scene dataset with 10096 images,and cumulative matching scores of top 2%is 91.5%,which outperforms most existing retrieval algorithms just performing one round of ranking.2)An improved manifold ranking based shoeprint retrieval algorithm is proposed.The existing crime scene shoeprint retrieval algorithms based on manifold ranking cannot describe the relationship between the partial shoeprints well.In order to improve this problem,this thesis proposes an improved manifold ranking algorithm for crime scene shoeprint retrieval.In addition,this thesis also gives a transformation function,through which the manifold ranking algorithm can achieve better performance in shoeprint retrieval.The algorithm indirectly measures the similarity between the two dataset images by the difference of similarity scores with the query respectively,which has a good effect on the crime scene shoeprints with low quality.The algorithm achieves a 1.8%improvement in the top 2%of the existing manifold ranking method on the crime scene shoeprints dataset,which proves the effectiveness of the algorithm.At the same time,the proposed manifold based algorithm can use the score computed by the proposed local semantic patch based algorithm as an input,and get a better reranking performance.The algorithm has also been tested on the reference shoeprints dataset,and does not cause a decline in retrieval performance,which can also prove the robustness of the algorithm.