ELFA-LFIQE: Epipolar plane image low-level features-aware light field image quality evaluator
Ping Zhao, Xiaoming Chen, Yuk Ying Chung, Haisheng Li · Computers & Electrical Engineering · 2025
Light field image quality measurement (LF-IQM) is a fundamental vision-based measurement task throughout the lifecycle of light field image processing . The conventional statistics-based LF-IQM methods heavily rely on prior knowledge while enduring high-complexity computation. Some recently proposed deep learning-based methods, although showing superiority over their statistics-based counterparts in complexity and training efficiency, suffer performance decline on low-resolution light field image datasets. In this paper, we propose a novel Epipolar plane image-based, Low-level Features-Aware Light Field Image Quality Evaluator (ELFA-LFIQE), which incorporates an innovative multi-task learning mechanism. During the training process of ELFA-LFIQE, low-level features are extracted from the earlier stages of convolutional neural networks (CNN) and fed to local multi-tasks at fully connected (FC) layers. The outcome of these local multi-tasks is weighted and summed to that of the global multi-task at the final FC layers to form the overall outcome. This multi-scale multi-task learning (MSMT) model enables a more holistic and accurate light field image quality assessment . To further improve the training efficiency, we propose a separable auxiliary task learning (SATL) mechanism that can significantly improve the generalization of the learning model. Another main contribution of this paper is that we propose a unified epipolar plane image (EPI) patch size (256 * 8) and a uniform network design that eliminate the needs of manual selection of EPI patches while maintaining training efficiency and accuracy, which we believe will foster more research into the EPI-based learning models for LF-IQM tasks. The extensive experiments show that our proposed model outperforms the state-of-the-art LF-IQM models on mainstream light field image datasets. We will open source our code, available on GitHub: https://github.com/ping-zhao-yz/elfa-lfiqe .