An Unsupervised Stereoscopic Image Quality Prediction Model Using Perceptual and Statistical Features of Scene Attributes

Abstract

Stereoscopic images (S3D), as an advanced visual multimedia format, are gaining popularity among consumers and researchers due to their ability to provide immersive and realistic experiences compared to conventional two-dimensional (2D) content. However, similar to the 2D imaging pipeline, S3D images are susceptible to quality degradation and perceptual losses during acquisition, encoding, transmission, and display systems. Therefore, automatic quality assessment (QA) of S3D images without pristine S3D features and training/testing mechanisms is of utmost importance. In this work, we propose an unsupervised no-reference (NR) S3D image quality assessment (IQA) model using a cohesive color map generated from the individual views of an S3D image. The quality-aware representations of the cohesive map are computed by partitioning it into non-overlapping blocks, followed by applying singular value decomposition to each block and utilizing its singular values and eigenvector matrices to capture the distortion perturbations of S3D images. Furthermore, the quality score of the cohesive map is computed by averaging the harmonic distance between the multivariate Gaussian parameter of pristine images and the corresponding block-level features of a test S3D image. Finally, a weighted parametric pooling function is employed to fuse the cohesive map quality score with spatial quality scores derived from the PIQE algorithm to obtain the final perceptual quality score of a test S3D image. The superiority of our proposed S3D quality assessment model has been demonstrated through extensive experiments involving sixteen off-the-shelf models across four benchmark S3D IQA datasets. The proposed model outperformed three opinion-unaware 2D NR IQA models and three S3D NR IQA approaches and also delivered competitive performance against ten opinion-aware S3D NR IQA models. Notably, this performance was achieved without relying on any regression or deep learning modules, highlighting the model’s robustness in estimating the perceptual quality of S3D images. The source code for the proposed opinion-unaware model can be accessed via the following: Google Drive link. © 2025 Elsevier B.V., All rights reserved.