Research On Sparse Representation Based Shoeprint Retrieval Algorithm

Shoeprints are often found at crime scenes and provide important evidences for criminal investigations.Because of the existence of repeated offences,effective retrieval of crime scene shoeprints not only assists the criminal investigation officers in linking and corporating cases,also can help them identify potential suspects.It is one of the most urgent tasks to retrieve the shoeprints efficiently.The existing shoeprint image retrieval algorithms mainly consider the correlation between the query image and images in the dataset,and they ignore the effect of same pattern shoeprint images on the retrieval results and the fact that features of the same class images are redundant in the dataset.Sparse representation uses less data to represent the image,and describes image features in a detailed way.Therefore,a shoeprint image retrieval algorithm based on sparse representation is proposed to improve the performance of the shoeprint image retrieval algorithm in this thesis.The main works of this thesis are as follows:1)A shoeprint image retrieval algorithm based on DCT domain is proposed.A shoeprint image retrieval algorithm based on DCT domain according to the characteristics of shoeprint image is proposed,and it makes the retrieval results more correlated with human opinions than the existing algorithms by shoeprint DCT domain feature and Fourier-Mellin feature based ranking.The experimental results show that the proposed shoeprint image retrieval algorithm has a better performance,and the recall rate of the top 10 and top 20 reach to 93.9%and 96.68%on the data set which contains 9294 shoeprint images.2)A shoeprint image retrieval algorithm based on multiple samples representation is proposed.The proposed algorithm includes the construction of multiple samples joint dictionary and the reconstruction of sparse coefficient with different weight.The algorithm expands the dataset automatically,which not only can improve the retrieval performance but also enhance the generalization ability of the algorithm.K-Means is used to cluster each class of shoeprint images in the expanded dataset,and the cluster centers of each class are the atoms of the dictionary.The sparse coefficients of the images are obtained by the orthogonal matching pursuit algorithm,and reconstruct the sparse coefficient using the proportion of pattern in the shoeprint.Experimental results on shoeprint image data set shows:The main performance indicators are better than the existing typical algorithms.The algorithm's recall rate of the top 10 and top 30 reach to 93.88%and 97.21%on the data set which contains 9294 shoeprint images.3)A shoeprint image retrieval algorithm based on Stacked Auto-Encoders and KNN is proposed.For an image can be represented by learning a multi-layer nonlinear network structure,a shoeprint retrieval algorithm based on Stacked Auto-Encoders and K-Nearest Neighbor is proposed.The algorithm ensures the accuracy of the retrieval precision,and is helpful for real-time processing of images.The algorithm's recall rate of the top 10 and top 20 reach to 87.9%and 94.4%on the data set which contains 9294 shoeprint images.