| The shoeprint image is one of the commonest traces in crime scenes, and it is a vital evidence to expose and verify crime. So far, conventional hand-crafted features have been explored by the shoeprint classification or retrieval algorithms to improve the performance and the designing of these features needs lots of engineering technologies and specialized knowledge. To some extent, it limits the progress of classification and retrieval algorithms. In recent years, Convolutional Neural Networks (CNN) which stands out from the field of image classification can learn a robust feature by imitating brain working mode. The process of learning feature through CNN doesn’t require too much specialized knowledge, and it can not only alleviate burdens of researchers, but also can improve the performance of the classification algorithms. Therefore, a shoeprint classification algorithm based on CNN is proposed in this thesis, which is expected to get robust feature through learning not by hand and show high classification accuracy.In practice, the crime scene shoeprints are always divided into many categroies and some categories contains a few samples. Moreover, the patterns of the shoeprints which belong to the same category are in large diversity. Thus, if the CNN model is trained by raw shoeprints directly, it will show poor performance and be hard to converge. The thesis studies shoeprint classification using CNN model in terms of training dataset and training efficiency and provides corresponding solutions, which can improve the classification accuray. And the main works are as follows:1) The thesis firstly introduces the structure, idea, working principle of CNN, some public available datasets and the corresponding models. Then the thesis analyzes the differences between the dataset used in the thesis and open datasets. Besides the thesis mentions integrated framework of the CNN model used to shoeprint classification.2) The thesis analyzes the possible difficulties in training CNN using small sample set. And then it proposes training methods using small sample set in terms of training dataset and training efficiency. The first one is about data augmentation and sample number selection; the next one is training acceleration.3) There are similar feature maps in the model, i.e., there are redundant connections. The thesis provides an optimization method through remove those extra connections. The optimization method not only accelerates the speed of training but also improves the performance of the model.The thesis combines the above-mentioned methods to classify shoeprints, and the model shows high performance. Compared to conventional CNN model and the method used hand-crafted feature, the error rate of the method provided in the thesis has reduced 6.57% and 2.07% respectively, and the classification accuracy achieves 97.57%. |
