Feature descriptors and learning approaches for image retrieval

Abstract

Content-Based Image Retrieval (CBIR) is a framework that locates, retrieves and displays images alike to one given as a query, using a set of features and image descriptors. In addition, suitable querying, matching, indexing and searching techniques are required. Content Based Image Retrieval (CBIR) has wide range of applications like medicine, digital libraries, satellite imaging, etc. Nowadays, the growth rate of the digital image databases is very high, which in turn require CBIR systems to handle large image databases efficiently. The overall performance of a CBIR system highly depends on the methods used for feature extraction and the function equipped for finding the image similarity. The feature extraction might be a global or local phase. Global (low level) phase extracts the features from the whole image where the local phase makes use of only the sub-regions of the image. In general, the low level feature includes color, shape, texture and spatial feature vectors. Two novel feature extraction methods based on color histogram and texture are proposed in this research work towards improving the image retrieval performance. After feature extraction, similarity measures are conducted to compare the query image against the images in the search database. Based on the similarity comparisons, the database images are indexed to find the closest results. These indexing could be refined iteratively by receiving the Relevance Feedback (RF) from the user. According to the feedback, the query and the similarity measures is updated at every iteration, so that the target image could be reached shortly. newline