Research On Feedback Based Crime Scene Shoeprints Retrieval Algorithm

Shoeprints play an enormous role to detect criminal cases. Experts often match shoeprints that are found at different crime scenes in order to detect series of cases. The most challenging task for a forensic examiner is to work with highly degraded footware marks and match them to the most similar shoeprint available in the database. In this paper, the crime scene shoeprints retrieval algorithm based on feedback that can be used to help the experts find the most matching image accurately is firstly proposed, and the retrieval results conform to the subjective judgment of experts. The contributions are listed as following:1) A crime scene shoeprints coarse retrieval algorithm is proposed(1) Feature extraction algorithm based on Wavelet-Fourier transformA hybrid wavelet transform and Fourier-Mellin transform based method is proposed to reduce the effect of noises, breaks and partial images. The proposed method combines the multiscale analysis characteristic of wavelet transform and the translation and rotation invariance of Fourier-Mellin transform.(2) Matching strategy based on local and global featureMatching strategy based on local and global feature is proposed to improve the accuracy of the partial images matching. The image matching strategy is based on correlation coefficients of each partion. For the partial images (i.e regions with lower confidence), similarities are estimated based on computed similarities of regions with higher confidence values. The final score is the weighted sum of the image partion similarity scores.The performance of coarse retrieval algorithm is evaluated by CMC, and the cumulative match scores of the first0.1percent, first0.2percent and first2percent are45.24,64.09and87.5respectively.2) The crime scene shoeprints second retrieval algorithm based on SVM is proposedFor the unfavorable retrieval results of the partial images and shoeprints with severe noises, this thesis proposes a hybrid feedback and support vector machine based method which extracts the common features and differences features in order to improve the retrieval accuracy. The distance of features is used to measure the similarity of images according to the classification results of each partion. The final score is the weighted sum of each partion distance based on partion confidence and classification results. The shoeprint images are sorted according to the classification results and final scores. The cumulative match scores of second retrieve are increased by7percent on average compared with the coarse retrieval results.Experiment results on the database composed of9,592crime scene shoeprints,72test images and432generated geometry distortion versions of test images show that the final cumulative match scores of the first0.1percent, first0.2percent and first2percent are54.56,70.24and87.5respectively. The proposed crime scene shoeprints retrieval algorithm based on feedback is robust to slight geometry distortions and interferences such as breaks and small holes, and noises, etc. This algorithm can also deal with partial prints on different levels.