Development of a content-based image retrieval system

In this thesis, a Content-Based Image Retrieval (CBIR) system is presented that supports querying with respect to color and texture low-level features. The fundamental idea is to generate automatically image descriptors by analyzing the image content. The focus will be on computing global similarity between images. Query is made upon images of homogeneous color/texture that do not require segmentation. The selected images domain is fashion and interior design.The underlying techniques are based on the adoption of Gray Level Cooccurrence Matrix (GLCM)and correlogram (correlation histogram)as statistical approaches to texture analysis.In addition,cumulative histogram and moments are utilized in color analysis. These techniques are applied in separated and combined manners.Each image is represented by features vector(s) in the features space.These vectors are indexed using an iterative clustering algorithm called Hierarchical Agglomerative Clustering (HAC) which provides easy-to-index data structures as well as faster query execution facilities. The degree of similarity between images is defined by the distance in the features space.Given a query image, the system first extracts its features vector, and then
compares this vector with those of the images pointed along the index structure using wide or narrow search algorithms. In this way, the matched images could be ranked and put into group according to the distance of their features vectors to the query one. This ranked group is considered as the query result.During the evaluation process a comparison study is made between different applied retrieval schemes. Cumulative histogram proved to be the best according to the selected domain, both as a separated retrieval scheme or when it is combined with GLCM or correlogram, respectively.The conducted experimental evaluation showed that the clustering based indexing algorithm offers high retrieval accuracy with a considerable reduction in the number of required similarity comparisons. Search efficiency is improved due to the fact that the query image is not compared exhaustively with all the images listed in the database.