Cloud Type Classification of Total-Sky Images Using Duplex Norm-Bounded Sparse Coding

Cloud type classification plays an essential role in ground-based cloud observation. However, it is a challenge to accurately identify the cloud categories that involve a large variety of structural patterns and visual appearances. Image representation and classifier are the crucial factors for cloud classification, though they are individually investigated in the literature. This paper proposes a new cloud type classification method using duplex norm-bounded sparse coding (DNSC), which designs image representation and classifier under the same framework, i.e., norm-bounded sparse coding (NSC). NSC is not only used to encode local descriptors, but also well explored to develop an effective classifier. NSC takes both locality and sparseness into account, and it can be benefit to capture discriminative patterns for image representation and have discriminative power for classifier. Furthermore, NSC has closed-form solution and can be computed efficiently. More specifically, DNSC first extracts local descriptor from an input cloud image, and then DNSC forms a holistic representation leveraging NSC and max-pooling strategy. Finally, a classifier is built on the holistic representation using NSC. The proposed DNSC is evaluated on the total-sky cloud image set, and the experimental results demonstrate that DNSC outperforms the state-of-the-art methods and its accuracy increases by about $7\%$ compared with baselines. In addition, the categorywise performance improvement is particularly pronounced over the complex categories, such as Cirriform and Mixed cloudiness.