Research On Semantic-Based Image Classification And Retrieval

Semantic based image classification and retrieval is being a academic hotspot in Computer Science field, whose main object is to classify and retrieve images precisely and efficiently. Some key problems are studied in this thesis, and some results are obtained:An approach to construct a color semantic feature to map the low-level feature to semantically meaningful categories is proposed. Aiming at overcome the semantic gap between low-level features and high-level semantic concept of image and to improve the precision of image classification. Color feature was extracted first and then a semantic network which contained a set of color concepts and objects related was built. Finally a color semantic 3-tuple was constructed. After classification, the experiment result demonstrated that the proposed method can not only get better result in image classification, but also was robust in different classifiers.A new model of visual perception - VAWO (Visual Attention Weight Order) is proposed for modeling the feature. The VAWO of every image region is calculated by image global features combined with image segmentation. Then the combined features are used for image classification using some machine learning algorithm. Experiments show that using this method to construct the feature vector in image retrieval, you can not rely on any external manual operation to obtain good retrieval accuracy, and the performance of the independence of the model is excellent.A image retrieval prototype system in both desktop version and web version is developed using java language. Both of them are running well in both Linux and Windows.To conclude, two approaches are proposed for improving the classification accuracy. A prototype system is developed based on the research for demonstrating the experiment result. Finally, the experiment result presents the significance and worthiness of the research.