Large-Scale Classification of Cargo Images Using Ensemble of Exemplar-SVMs

摘要

This paper develops a large-scale classification algorithm for cargo X-ray images using ensemble of exemplar-SVMs. Large-scale or fine-grained classification is very helpful for customs to improve the inspection efficiency and liberate their inspectors. However, big intra-class variation accompanied with small inter-class variation of cargo images makes it almost impossible to classify them using traditional class-SVM between classes. But those typical images with salient and representative features of some classes can be easily distinguished from others. Inspired by the idea of the ensemble of exemplar-SVMs for object detection, we develop a classification method using the ensemble of exemplar-SVMs of typical image patches. Firstly typical image patches are defined and the method of extracting them is discussed. Then for each of the typical image patches, a linear SVM is trained using itself as the positive sample and all others as the negative samples. In the classification step, fast detection method based on WTA-hash is used. Images are firstly classified to typical patches and then classified to the category of the corresponding typical image patches. A semantic tree built according to the HS code is used to trade off specificity for accuracy.