Unsupervised Content-Based Image Retrieval Technique Using Global and Local Features

Content-based image retrieval is an application of retrieving similar images of a query image. In this paper, global and local features are used to describe the content of an image. Color, texture and shape features are used as global features and SURF, FAST and BRISK are used as local features to retrieve similar images from the database. Different similarity checking methods are also tested to find the best approach. Combined approach using global features (obtained by using modified gray-level co-occurrence matrix) and local features (obtained by using SURF and color moments) produce better accuracy than some existing methods. Here, local features are obtained around the blob points detected by SURF detector. SURF descriptor and color moments which are calculated within a region of size 6s (s is the scale of the image at which the blob point is detected) around the detected blob points are used for extracting local features. Color moments are used because SURF only works on grayscale images. Thus, local color information is achieved. This combination of global and local features result in better performance in term of accuracy compared to some current methods.