| A shoeprint is one of the traces of high retention rate on crime scenes.It plays a very important role in shrinking the scope of the suspect and analyzing joint cases.Shoeprint classification can classify shoeprints in terms of the features reflected in the images.Semantic concepts can improve the shoeprint classification performance effectively.However,the current semantics based classification algorithms are not perfect enough,because they ignored two relationships.The first one is the relationship between the primitives’ categories and their spatial location characteristics,and the second one is the relationship between the primitives’categories and the shoeprint categories.So we proposed a unique attribute and local semantics based shoeprint classification algorithm.It consists of following four parts:1)A unique attribute and local semantics based shoeprint classification framework is proposed.In terms of the characteristics of shoeprints and the current shoeprint classification models,we proposed an algorithm framework of shoeprint classification.The framework mainly consists of three parts:the local semantic representation combined with unique attribute,the multi-scale primitive substitution algorithm to obtain primitive distribution characteristics and the decision-level ensemble classification algorithm based on mixed weight.2)A unique attribute and local semantics based shoeprint representation method is proposed.The local semantic representation algorithm of shoeprint is given by taking into account the relationship between the primitive categories and their spatial location characteristics.And according to the relationship between the primitive categories and the shoeprint categories,a unique attribute dictionary is constructed.Based on.analyzing the role of unique attribute characteristics in the classification of shoeprints under different conditions,we proposed a special word feature and analyzed its classification performance.Finally,we proposed a local semantic representation combined with unique attribute characteristics.3)A spatial relation description algorithm based on multi-scale primitive substitutions is proposed.We replaced all primitives with multiple scale primitives and fused the features to obtain the representation of the spatial relationship.We also analyzed the influence of the scale and the number of alternative primitives on spatial relationship description and gave a reasonable combination of primitive substitution scales.4)Mixed weights based decision-level ensemble classification algorithm is proposed.In this thesis,a decision-level ensemble classifier based on mixed weights is used to make decision-level fusion of multiple shoeprint features.It used mixed weights to obtain high-confidence classification results.The similarity calculation method,the appropriate classifieation decision and the reasonable weights selection method are proposed,which are more suitable for the characteristics of shoeprints.The gallery set consists of 5491 suspect shoeprints from 3,500 classes,and the probe set consists of 1143 shoeprint images from 1143 classes.The accuracy rate of the proposed algorithm has reached 94.40%,and the classification results have good subjective and objective consistency.The classification accuracy rate is further improved than the current semantic-based shoeprint classification algorithm whose classification accuracy rate is 92.90%. |
