Full Reference Stereoscopic Objective Quality Assessment using Lightweight Machine Learning

Abstract

Decades of research on Image Quality Assessment (IQA) have promoted the creation of a variety of objective quality metrics that strongly correlate to subjective image quality. However, challenges remain when considering quality assessment of 3D/stereo images. Multiple objective quality metrics for 3D images were designed by extending the well-known 2D metrics. As a result, these solutions tend to present weaknesses under 3D-specific artifacts. Recent works demonstrate the effectiveness of machine-learning techniques in the design of 3D quality metrics. Although effective, some machine learning-based solutions may lead to high computational effort and restrict its adoption in low-latency lightweight systems/applications. This paper presents a study on full-reference stereoscopic objective quality assessment considering lightweight machine learning. We evaluated four different decision tree-based algorithms considering eight distinct sets of image features. The classifiers were trained using data from the Waterloo IVC 3D Image Quality Database to determine the subjective quality score measured using Mean Opinion Score (MOS). The results show that RandomForest generally obtains the best accuracy. Our study demonstrates the feasibility of decision tree-based solutions as an accurate and lightweight approach for 3D image quality assessment.