Research On Open Set Shoeprint Classification Algorithm

Shoeprints are very important evidences for criminal investigations.In order to make shoeprints play important roles in case detection,forensic officers need to establish a shoeprint database.When a new shoeprint is automatically added to the database,the new shoeprint may belong to a known class or not belong to any classes in the database.The existing open set classification algorithms can't get good performance when they are directly used to classify shoeprints,therefore,it is necessary to design a shoeprint open set classification algorithm.The main work of this thesis is as follows:1)An algorithm named CSoftmax based on confidence is proposedFor the limitations of current Softmax classification algorithm existing when applied to open set scenarios,we intend to improve the accuracy of the algorithm through increasing the probability distribution difference between known categories and new categories by adding confidence.Adding confidence can reduce the probability overlap regions between the two categories.The experiment on shoeprint dataset which has obvious margins shows that proposed algorithm has better performance,and its AUC can reach 76.33 percent.2)An algorithm named DKNFST based on distance comparison information is proposedAccording to the distance characteristic between known and new categories in zero space,this proposed algorithm increases the distribution difference between known and new categories by considering the information included not only the nearest class but the most remote two classes.The experiment on shoeprint dataset which has obvious margins shows that its AUC can only reach 74.16 percent.3)An algorithm named MKNFST based on manifold consistency is proposedBased on the manifold consistency between the raw high dimensional feature space and the transformed zero space of the sample to be detected,the proposed algorithm constructs a classifier by fusing similarity values in two spaces,and it can further improve the accuracy of classification.The experiment on real shoeprint dataset which has obvious margins shows that its AUC can reach 83.77 percent.