Enhancement Of Feature Discriminant Ability With Synergistic Solution For Scene Categorization

The image scene classification is an important part of the high-level semantic image understanding, aiming at the semantic annotation process by analyzing the images on the overall statistics and associated features. The correct scene classification can find the similarity of the same scene category and the distinction of different scene categories, and it can provide the context relations between scenes and objectives in order to get more correct recognition results of target in the scene. Nodes of synergetic model contain natural statistical dependencies, reflecting the optimal division of synergetics, so it can be used for scene classification.In this thesis, Synergetic Network is build based on scene visual information description under the framework of the optimal solution of synergetics. It studies scene features separability and prototype vector expression and gets the steady state solution through scene features evolution. The main work of this thesis includes:1. Comparison of two global features:PHOG features and gist features. They can form prototype vector expression and input feature vector expression. Manifold learning methods can effectively reduce the feature dimension to obtain the order parameter expression of synergetic network.2. The extraction and application of salient features. Salient features are obtained by maximum symmetric surround saliency detection algorithm and they can be used to complete the image segmentation and classification effectively.3. Some images cannot obtain the proper labels from classfied decision of support vector machine but rather transferring labels through the evolution of order parameter evolution equation in synergetic network, which lead to more accuracy for scene categorization. The experiments verify the validity of this method.