Scene image recognition with multi level resolution semantic modeling

In this paper, we propose a multi-level resolution semantic modeling for automatic scene recognition. The basic idea of the semantic modeling is to classify local image regions into semantic concept classes such as water, sunset, or sky, and use occurrence frequency of local region´s semantic concepts for global image representation. However, how to decide size of the local image regions is a trial problem. The optimized region size would be dynamically changing for different scene or concept types. Therefore, this paper proposed a dynamical region size (Multi-level resolution) of local image regions for semantic concept model, and fusion the probabilities to scene types of several resolutions for final recognition of a scene image. Experimental results show that the recognition rate using our proposed algorithm is much better than that using the conventional semantic modeling method for scene recognition.