Query-by-example HDR image retrieval based on CNN

Abstract

Due to the expension of High Dynamic Range (HDR) imaging applications into various aspects of daily life, an efficient retrieval system, tailored to this type of data, has become a pressing challenge. In this paper, the reliability of Convolutional Neural Networks (CNN) descriptor and its investigation for HDR image retrieval are studied. The main idea consists in exploring the use of CNN to determine HDR image descriptor. Specifically, a Perceptually Uniform (PU) encoding is initially applied to the HDR content to map the luminance values in a perceptually uniform scale. Afterward, the CNN features, using Fully Connected (FC) layer activation, are extracted and classified by applying the Support Vector Machines (SVM) algorithm. Experimental evaluation demonstrates that the CNN descriptor, using the VGG19 network, achieves satisfactory results for describing HDR images on public available datasets such as PascalVoc2007, Cifar-10 and Wang. The experimental results also show that the features, after a PU processing, are more descriptive than those directly extracted from HDR contents. Finally, we show the superior performance of the proposed method against a recent state-of-the-art technique.