Purchase this article with an account.
Shuichi Takahashi, Takafumi Morifuji, Masami Ogata, Anthony Norcia; EEG-based classification of images as HDR versus non-HDR using Steady-State Visual Evoked Potential. Journal of Vision 2017;17(10):773. doi: 10.1167/17.10.773.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Recently, high dynamic range (HDR) images and HDR-ready displays have become available. Compared to the conventional standard dynamic range (SDR) images, HDR images have higher peak luminance and contrast, which leads to more faithful reproduction of actual scenes. Although subjective assessment methods such as the mean opinion score are often applied in order to classify captured or processed images as HDR or non-HDR, results based on these methods are biased by raters' experience and expertise. To address this issue, research on image-quality assessment using electroencephalography is widely performed. Here we determined whether SSVEPs can be to classify test images into HDR or non-HDR content. In this study SSVEPs for seven subjects were recorded during the presentation of five test stimuli. The stimuli consisted of HDR, SDR, and blended images of the HDR and the SDR images. Each SDR image was processed so that its average luminance was identical to the corresponding HDR image. They were presented on a professional HDR monitor at the frequency of 3 Hz. In addition, 5-scale mean opinion scores were collected for the stimuli before or after the EEG recordings. The results showed that the response amplitudes of the SSVEPs at the stimulus frequency were highly correlated to the mean opinion scores for most of the stimuli. We also constructed an HDR/non-HDR classification model based on a support vector machine (SVM) with a non-linear kernel. In the machine learning, standardized values of the amplitudes at channels over the visual cortex and those of the mean opinion scores were selected for student and teacher datasets, respectively. A data resampling technique was also applied to the skewed datasets. The classification accuracy of this subject- and image-independent model was 0.73.
Meeting abstract presented at VSS 2017
This PDF is available to Subscribers Only