Abstract
Effectively assessing the realism and naturalness of images in virtual (VR) and augmented (AR) reality applications requires Full Reference Image Quality Assessment (FR-IQA) metrics that closely align with human perception. Deep learning-based IQAs that are trained on human-labeled data have recently shown promise in generic computer vision tasks. However, their performance decreases in applications where perfect matches between the reference and the distorted images should not be expected, or whenever distortion patterns are restricted to specific domains. Tackling this issue necessitates training a task-specific neural network, yet generating human-labeled FR-IQAs is costly, and deep learning typically demands substantial labeled data. To address these challenges, we developed ConIQA, a deep learning-based IQA that leverages consistency training and a novel data augmentation method to learn from both labeled and unlabeled data. This makes ConIQA well-suited for contexts with scarce labeled data. To validate ConIQA, we considered the example application of Computer-Generated Holography (CGH) where specific artifacts such as ringing, speckle, and quantization errors routinely occur, yet are not explicitly accounted for by existing IQAs. We developed a new dataset, HQA1k, that comprises 1000 natural images each paired with an image rendered using various popular CGH algorithms, and quality-rated by thirteen human participants. Our results show that ConIQA achieves superior Pearson (0.98), Spearman (0.965), and Kendall's tau (0.86) correlations over fifteen FR-IQA metrics by up to 5%, showcasing significant improvements in aligning with human perception on the HQA1k dataset.