A hybrid generative-discriminative learning algorithm for image recognition

Abstract

Feature representation is usually a key point in image recognition. The recognition performance can be potentially improved if the data distribution information is exploited. In this paper, we propose an image recognition approach based on generative score space. Specifically, we first leverage probabilistic latent semantic analysis (pLSA) to model the distribution of images. Then, we derive the mid-level feature from the model in a generative feature learning manner. At last, the derived feature is embedded into a discriminative classifier for image recognition. The advantages of our proposed approach are two folds. First, the probabilistic generative modeling allows us exploiting information hidden in data and has good adaptation to data distribution. Second, the discriminative learning process can utilize the information of label effectively. To confirm the effectiveness of our method, we perform image recognition on three datasets. The results demonstrate its advantages.