Image classification technology based on mining of frequent item sets

We propose a novel method to detect frequent and distinctive feature configuration on a class instance. Each neighborhood of a local feature is described by a list of nonzero indices, and generates a transaction. An efficient mining of frequent item sets algorithm is used to automatically find spatial configurations of local features occurring frequently on a class instance. These mined spatial configurations can be used as special words, incorporate into bag of features classification model. Through evaluation on PASCAL 2007 Visual Recognition Challenge dataset set, the test results show that this mining algorithm is computationally efficient and allows to process large training sets rapidly. Moreover, the mined feature configurations have higher discriminative power compare to individual features.