Evaluating the Impact of Perceptually Uniform Encoding on the Performance of Feature Descriptors for HDR Image Retrieval

Recent advances in High Dynamic Range (HDR) imaging have led to various research studies. The features extraction of this kind of images is considered as one of the crucial tasks to several multimedia applications. In this paper, we look forward to testing and studying the impact of the Perceptually Uniform (PU) encoding on HDR feature descriptors. Specifically, we propose to measure and compare the effectiveness of Convolutional Neural Networks (CNN) and Scale Invariant Feature Transform (SIFT) descriptors in the context of HDR image retrieval. Experimental evaluations show that the PU encoding leads to substantial performance improvements over the linear HDR content for the two above-mentioned descriptors. These experiments also demonstrate that the CNN features are more effective than SIFT ones according to the mean Average Precision (mAP) scores. Furthermore, SIFT descriptor is inappropriate for linear HDR content.