Exploiting Class Hierarchies for Large-Scale Scene Classification Using Hybrid Discriminative Approach

Humans seamlessly perceive a massive amount of information while observing a scene. Though humans recognize real-world scenes easily and accurately but its not the same for computers due to scene images variability, ambiguity, and diverse illumination and scale conditions. Scene classification is a fundamental problem which provides contextual information to guide other processes, such as browsing, content-based image retrieval and object recognition. A baseline model based on traditional bag of words model is built to better evaluate the proposed solution. Model based on the idea of fine to coarse category mappings is proposed, whose information is combined with the fusion of feature descriptors resulting in a single feature representation. This additional information enhances performance by exploiting hierarchical relationship among the scene categories. Effectiveness of the proposed approach is validated using different evaluation metrics. Proposed model performs considerably better compared to the given baseline as well as several state-of-the-art methods. Proposed framework ensures appropriate balance between time and accuracy of the model