Content-Based Image Retrieval using Deep Learning.

摘要

A content-based image retrieval (CBIR) system works on the low-level visual features of a user input query image, which makes it difficult for the users to formulate the query and also does not give satisfactory retrieval results. In the past image annotation was proposed as the best possible system for CBIR which works on the principle of automatically assigning keywords to images that help image retrieval users to query images based on these keywords. Image annotation is often regarded as the problem of image classification where the images are represented by some low-level features and the mapping between low-level features and high-level concepts (class labels) is done by some supervised learning algorithms. In a CBIR system learning of effective feature representations and similarity measures is very important for the retrieval performance. Semantic gap has been the key challenge in the past for this problem. A semantic gap exists between low-level image pixels captured by machines and the high-level semantics perceived by humans. Machine learning has been exploited to bridge this gap in the long term. The recent successes of deep learning techniques especially Convolutional Neural Networks (CNN) in solving computer vision applications has inspired me to work on this thesis so as to solve the problem of CBIR using a dataset of annotated images.