Research On Fine Grained Shoeprint Image Retrieval

Shoeprint is one of the criminal evidences with high retention rate at the scene,which plays a vital role in the criminal investigation.By comparing the shoeprints collected at the crime scene with the shoeprints in the dataset,criminal investigators can link several cases and increase the probability of detection.As the number of shoeprints with similar textures proliferates,the differences between the shoeprints in the database are decreasing,which is a challenge for the shoeprint image retrieval task.Therefore,it is of great theoretical and practical significance to design a retrieval algorithm for crime scene shoeprints.Existing shoeprints retrieval algorithms mainly focus on extracting coarse grained global features,ignoring fine-grained information in the image,such as the semantic regions with distinctive signs and the spatial location distribution in the shoeprints.In this thesis,a finegrained shoeprints retrieval algorithm is proposed,and the main works is as follows:1)A distinguishability semantic region automatic selection algorithm is designed.The algorithm can automatically select multiple distinguishable semantic regions.The regions are clear and representative of the characteristics of the shoeprint,highlighting the fine-grained information of the shoeprint image.It also effectively avoids the problem of duality caused by manual selection of semantic regions.By training semantic filters on the semantic regions,the similarity score between the shoeprint image to be retrieved and the dataset image is calculated.And the feature similarity score obtained based on wavelet Fourier Mellin features is fused to obtain the final retrieval score.Experimental results are conducted on three shoeprint datasets and good results are obtained.On the crime scene dataset MUES-SR10KS2 S,the cumulative matching scores between the automatically and the manually selected semantic regions in the top 2% only differs by 3 percentage points.2)A similarity enhancement algorithm based on content and spatial layout relationship is proposed.In the process of calculating the similarity score between semantic regions and the dataset,the algorithm not only considers the similarity between features,but also takes into account the fine-grained information of the spatial layout distribution of the semantic regions.Experimental results are conducted on three shoeprint datasets and good results are obtained.On the crime scene dataset MUES-SR10KS2 S,the cumulative matching score of the top 2% is90.5%.3)A shoeprint retrieval algorithm based on graph representation and manifold ranking is proposed.The algorithm constructs a shoeprint graph based on multiple semantic regions to represent the shoeprint image,and calculates a similarity score between query image and dataset images.Fusing the enhanced similarity scores based on content and spatial layout relationships to obtain fine-grained scores between the query image and dataset images.The final retrieval result is output by combining the manifold ranking algorithm.Experimental results are conducted on three shoeprint datasets and good results are obtained.On the crime scene dataset MUES-SR10KS2 S,the cumulative matching score in the top 2% reached 91.7%.