Scene classification based on knowledge sharing and latent structural constraints

We address the issue when applying local discriminative information to construct the mid-level representing for scene classification. It is often caused by the confusion of local semantic similar when convolving local cues with input scene images. This problem makes mid-level semantic representation for images, not distinctive enough to classify those unseen data correctly. We first learn knowledge sharing for exploring distinctive local semantic information based on latent semantic analysis and label ranking. Then, we employ the latent structural constraints of scene configuration for the semantic prediction. Finally, we infer the scene semantic using Bayes rules. Extensive Experimental results have illustrated the effectiveness and efficient of our method for local semantic annotation and scene classification.