Randomized Probabilistic Latent Semantic Analysis for Scene Recognition

The concept of probabilistic Latent Semantic Analysis (pLSA) has gained much interest as a tool for feature transformation in image categorization and scene recognition scenarios. However, a major issue of this technique is overfitting. Therefore, we propose to use an ensemble of pLSA models which are trained using random fractions of the training data. We analyze empirically the influence of the degree of randomization and the size of the ensemble on the overall classification performance of a scene recognition task. A thoughtful evaluation shows the benefits of this approach compared to a single pLSA model.