Discriminative Probabilistic Latent Semantic Analysis with Application to Single Sample Face Recognition

Face recognition is still a challenging issue due to the presence of intrinsic complexity, external variations and number limitation of training samples. In this paper, a novel face recognition method based on probabilistic latent semantic analysis (pLSA) model is developed, which mainly contains two stages: bag-of-words features extraction and semantic representation learning. In the first stage, to extract more structure information, the region-specific dictionary strategy is employed, i.e., generating a dictionary for each region. The encoded and sum-pooled features of all regions are concatenated together. In the second stage, a discriminative pLSA (DpLSA) model is presented, which initializes the word-topic distribution $$P(w|z_k)$$P(w|zk) by the center point of the training data from category k. As a result, the problem of how to choose appropriate number of topics in classical topic model is alleviated, and the training process of DpLSA is very fast only requiring few iterations. Moreover, the discovered topic-document distribution $$P\left( z|d\right) $$Pz|d is discriminative and semantic with the dominant topic entry corresponds to the category label of image d, which enables performing classification by $$P\left( z|d\right) $$Pz|d directly. Extensive experiments on four representative databases demonstrate that the proposed DpLSA is effective for face recognition under single training sample and possesses a certain degree of robustness to illumination, pose, as well as occlusion.