Automatic scene recognition for digital camera by semantic features

Accurate calibration is prerequisite for digital camera to get satisfactory images. However, various scene types need different camera calibration schemes. A few fixed scene modes on digital camera to facilitate the users have been proposed. These common scene modes (e.g. landscape, portrait, night scene, etc.) for daily use are optimized for specific scenes and photographic conditions. When selected, a scene mode can often give better results than shooting in fully automatic mode. In this paper, an approach for automatic recognition of scene types based on semantic features is presented. Latent Dirichlet Allocation (LDA) based topic model is adopted to generate semantic features from Scale Invariant Feature Transform (SIFT) image descriptors. Semantic features in this approach are not only a better dimensional representation for original image data, but also reports satisfactory classification performances on datasets of complex scenes, especially for small size training sets. Furthermore, as it is not possible for fixing all scene types beforehand in camera, our approach gives an option for the users to define new scenes through a cluster-based retraining method, only several new training examples are required. Experimental results show that the proposed approach is effective and flexible for automatic scene recognition in camera.