Research On Semantic Feature Based Shoeprint Classification Algorithm

Shoeprints are often found at crime scenes and provide valuable forensic evidences.They play very important roles in linking and incorporating cases analysis,and can narrow the scope of the searching suspects.Shoeprint image classification is an automatic process assigning an unknown shoeprint to the known category according to certain rules.It is mainly based on image understanding,which requires not only low level visual feature but also semantic concepts.Most of all,semantic concepts can effectively improve the performance of shoeprint classification.Therefore,we propose a shoeprint image classification algorithm based on semantic features.And the main works are as follows:1)A framework of shoeprint image classification algorithm based on semantics is proposed.The framework mainly includes three parts:semantic representation algorithm,spatial relationship description algorithm and nearest neighbor combined with decision tree classification algorithm.Experiments have been done in the suspect data set to verify the effectiveness of the proposed framework.2)Semantic representation algorithm considering semantic relevance is proposed.Based on the model of the traditional bag of visual words,combining the semantic characteristics of shoeprint images,a shoeprint image semantic vocabulary is firstly constructed.Then based on the idea of supervision and feedback,the semantic relations of the words in the semantic vocabulary are obtained.Finally,according to the semantic relations,the semantic representation of semantic relevance is realized.The experimental results show that the semantic feature of semantic relevance is more effective than the semantic feature without considering the relevance.3)Hierarchical spatial relation description algorithm is proposed.A shoeprint image is divided into different levels based on the anti-jamming ability of the element distribution.And different spatial relation description methods are used for different levels.Two kinds of spatial relation description methods are presented,and they are spatial relation description based on metric matrix and spatial relation description based on Wavelet-Fourier Transform respectively.The experimental results show that the proposed hierarchical spatial description algorithm can effectively describe the spatial relationship and provide a basis for classification.4)Decision tree classification algorithm combined with nearest neighbor is proposedThe tree structure of the decision tree is used to judge the credibility of the nearest neighbor classification results with different features,so as to obtain the classification results with high reliability.The algorithm retains the classification advantage of each feature,simplifies the training process of the classifier,and improves the classification accuracy through fusing two classification algorithms.The experimental results show that the proposed algorithm has better classification performance.In order to verify the classification performance of shoeprint image classification algorithm based on semantic,we construct a dataset composed of 3500 kinds of suspects shoeprint image,and the total shoeprint number is 7834.One thousand training images of one thousand different classes are randomly selected from the dataset,and 1143 test images of 1143 different classes are chosen from the dataset.The experimental results show that the classification accuracy has reached 92.9%,even though many mis-classified ones are well correlated with human opinions.