Topic-based label distribution learning to exploit label ambiguity for scene classification

One of the greatest challenges for scene classification is the lack of sufficient training samples. Label distribution learning (LDL) is proven to be effective in handling insufficient samples by exploiting label ambiguity. However, LDL has never been used in scene classification because the correlations among scene classes are unavailable, making it impossible to construct label distribution vectors for images. In this paper, we aim to transform LDL into scene classification. To this end, we introduce a probabilistic topic model (PTM) to capture label correlations, and propose a method termed as topic-based LDL (TB-LDL). By treating scene classes as documents in the PTM, the discovered topics indicate typical scene patterns, and class-topic distributions provide label measurements on multiple topics. For each topic, scenes with similar label measurements can be considered as neighbouring labels. The label distributions smooth image truth labels based on label correlations, which can formulate the label ambiguity of scene images. Training networks with the label distributions can prevent over-fitting and assist feature learning. Extensive experiments on two challenging datasets, namely the aerial image dataset (AID) and NWPU_RESISC45 (NR), demonstrate that our method is effective, especially when the amount of training data is limited.