Content-Based Image Retrieval Using Deep Learning

Abstract

In last few decades, digital images are growing with a rapid pace on and off the Internet, and given the volume of the images, the need of better storage, processing, and retrieval of images has raised. One of the retrieval techniques that is focus of this work is content-based image retrieval (CBIR) in which similar images are searched from a pool of images without manually annotating them; rather, in CBIR, other features of images that discriminate them from other images are used. By finding the better discriminative features of a collection of images, an efficient and generalized CBIR system can be built. The success and the efficiency of such a system depend on the choice of the features of images used to identify them. Generally, the images are stored at a very low level in pixels, but to get better results or in other words better features, we need to store these images at a very high level in order to reduce the semantic gap. There has been very extensive research on CBIR using the traditional methods of image processing. With the advancement of deep learning systems, deep learning can be used for large range of problems. In this work, we investigated the use of deep learning, more precisely auto-encoders, for the feature extraction and representation of images in CBIR, and we reached to the retrieval efficiency of ≈ 80%.