Discriminative learning of I2C distance for image classification

In order to improve image classification performance of the Image-To-Class (I2C) distance, a new distance learning method is proposed to improve the discrimination of I2C distance by learning the parameters in a regularized logistic regression framework in this paper. To generate a more discriminative I2C distance, we use the spatial layout of the individual features to further improve our learned distance. Meanwhile, a new kernel by making use of the output of the regression model can further enhance its complementary to the widely used bag-of-features based approaches. Experimental results can show that the proposed method can significantly outperform other I2C methods in several prevalent image datasets.