Utilizing Dictionary Learning and Machine Learning for Blind Quality Assessment of 3-D Images

Abstract

In recent years, perceptual objective quality assessment of 3-D images has become an intense focus of research. In this paper, we propose an efficient blind 3-D image quality assessment (IQA) metric that utilizes binocular vision-based dictionary learning (DL) and ${k}$ -nearest-neighbors (KNN)-based machine learning (ML) to more accurately align with human subjective judgments. More specifically, in the DL stage, histogram representations from the local patterns of simple and complex cells are concatenated to form basic feature vectors. Then, by using a collaborative representation algorithm, the learned binocular quality-aware features of the distorted 3-D image can be efficiently represented by a linear combination of only a few of these basic feature vectors. In the ML stage, we intuitively simulate the complex high-level behaviors of human perceptual activity with KNN-based ML, which transfers the weighted human subjective quality scores from the annotated 3-D images to the query 3-D image. Our results using three standard subject-rated 3-D-IQA databases confirm that the proposed metric consistently aligns with the subjective ratings and outperforms many representative blind metrics.