A novel strategy for classifying perceived video quality using electroencephalography signals. (June 2020)

• Record Type: Journal Article
• Authors: Chan, Kit Yan; Arndt, Sebastian; Engelke, Ulrich
• Is Part Of: Engineering applications of artificial intelligence. Volume 92(2020)
• Publication Date: 2020-06
• Subjects: Video quality -- Electroencephalography -- Quality of experience -- Affective computing -- Classification; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2020.103692
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: Video streaming through the Internet is abundant nowadays. While video quality is continuously demanded, monitoring users' quality of experience (QoE) is essential when watching video contents. QoE can be evaluated directly through subjective assessment which is the human ground truths; however, such assessment is generally expensive and time consuming, and cannot be implemented in real time. QoE can also be evaluated by video quality models; however, the evaluation is fully based on video contents but human physical states cannot be taken into account. To tackle the limitations, detection of a prominent electroencephalography (EEG) signal feature namely P300 correlated to QoE can be used, when users are viewing videos. P300 is a positive deflection pulse that appears around 300 ms after a significant video distortion appears. QoE can be indicated by P300 pulses. However, the captured EEG signal is generally contaminated with noise. Strong noise generates P300 although video carries no distortion. Hence, detections of P300 patterns are not accurate. In this paper, a double classifier consisting of a first and second classifier is proposed. The first classifier attempts to determine whether the captured EEG feature is abnormal or not, where the abnormal caption behaves opposite to the normal P300 characteristic when showing the distorted video. The second classifier is developed to perform classifications for either normal or abnormal features. We evaluate the performance of the proposed double classifier based on the EEG samples, which are captured when showing video stimuli to participants. The proposed classifier is implemented by the support vector machine and logistic regression, which are commonly used for detection of EEG patterns and are computationally much simpler than deep learning. The performance of the proposed classifier is compared to those of the single classifiers, which determine the QoE directly when the EEG signal is given. Cross-validations showed that generally more than 5% improvement can be achieved by the proposed double classifier. Statistical tests indicate that the proposed double classifier can generally obtain better classification rates than solely using the single classifier at a 97.5% confidence level.