Robust and distortion-aware visual representations are paramount for perceptual image quality assessment (IQA) and downstream visual understanding under real-world degradations. In this paper, we conduct a comprehensive analysis of state-of-the-art visual encoders, including CLIP, DINO, ConvNeXt, EfficientNet, and ResNet, under common distortions such as Gaussian blur, motion blur, compression artifacts, and Gaussian noise. We employ latent feature divergence, ANOVA-based effect size analysis, and dimension-wise mean absolute differences (MAD) to assess model robustness and sensitivity. Our results reveal how architectural choices, training objectives, and data diversity shape a model’s ability to encode distortions within its latent space. These findings bridge representation learning and perceptual quality modeling, offering new insights into the development of distortion-resilient encoders for IQA.

This paper introduces a novel graph-based framework for point cloud quality assessment (PCQA) that bridges the gap between objective metrics and perceptual quality. The framework constructs local graphs around key points identified through curvature analysis and adaptively incorporates spatial and color information to model local structures effectively. The proposed approach captures structural and visual characteristics by leveraging signal processing on graphs and extracting domain-specific features from the geometry, appearance, and spectral domains. Extensive evaluations using SJTU-PCQA and WPC, two publicly available datasets, demonstrate the superiority of the proposed framework, achieving state-of-the-art performance and setting new benchmarks among graph-based approaches for PCQA. Furthermore, the ablation study showcases the complementary benefits of integrating domain-specific features, highlighting the robustness and adaptability of the proposed method across various point cloud densities and qualities.

Super-resolution (SR) techniques play a pivotal role in enhancing the quality of low-resolution images, particularly for applications such as security and surveillance, where accurate license plate recognition is crucial. This study proposes a novel framework that combines pixel-based loss with embedding similarity learning to address the unique challenges of license plate super-resolution (LPSR). The introduced pixel and embedding consistency loss (PECL) integrates a Siamese network and applies contrastive loss to force embedding similarities to improve perceptual and structural fidelity. By effectively balancing pixel-wise accuracy with embedding-level consistency, the framework achieves superior alignment of fine-grained features between high-resolution (HR) and super-resolved (SR) license plates. Extensive experiments on the CCPD and PKU dataset validate the efficacy of the proposed framework, demonstrating consistent improvements over state-of-the-art methods in terms of PSNR, SSIM, LPIPS, and optical character recognition (OCR) accuracy. These results highlight the potential of embedding similarity learning to advance both perceptual quality and task-specific performance in extreme super-resolution scenarios.

This paper delves into the intricate field of perceptual quality assessment specifically tailored for 360-degree images, aiming to advance the precision of quality models. In contrast to conventional methodologies that associate different regions within images to mean opinion scores (MOS), our study introduces a paradigm shift. We propose a novel approach where the model is trained to predict multi-labels derived from subjective and objective measures, leveraging a sophisticated quality labeling framework designed to capture nuanced perceptual distinctions across diverse regions in panoramic content. This allows for a flexible and stable training process. In addition, a loss function taking into account the magnitude and direction of quality labels under a multi-label learning scheme is designed, using absolute and directional distances as loss functions. Experimental results underscore the limitations of relying solely on MOS as unique labels. The efficiency of our approach becomes evident through improved performance, showcasing its potential for advancing the precision of perceptual quality assessment models.

Super-resolution (SR) techniques play a crucial role in enhancing the quality of low-resolution images, with significant applications in fields such as security and surveillance, where license plate recognition is critical. This paper focuses on optimizing the super-resolution of license plates using embedding similarity learning. We proposed a novel framework that integrates a Siamese network with a super-resolution model to guide the SR model into enhancing the perceptual quality of reconstructed license plates. By leveraging embedding similarity through Contrastive loss, our approach ensures that the super-resolved images are perceptually and structurally closer to the original ones. The experiments on a synthetic dataset demonstrated that the proposed method outperforms traditional techniques that rely solely on pixel-based loss functions such as MSE. The introduction of embedding similarity loss significantly improves the PSNR and LPIPS metrics, in addition to the optical characters recognition rate.

In this study, we propose a methodology for estimating quality score distribution (QSD) for 360-IQA patch labeling. A collection of 2D-IQA models is used to generate a QSD for patches, inspired by how subjective quality ratings are gathered and handled. The proposed framework is first benchmarked on a subjectively annotated dataset, namely KonPatch-32k, in terms of patch quality classification. The best composition of QSD is then used to derive quality labels for patches sampled from 360-degree images. Furthermore, the quality labels are used in a multi-regression training strategy of CNN models. The ResNet-50 and EfficientNet-B5 are used to test the effectiveness of the proposed labeling framework on two publicly available 360-IQA datasets, namely OIQA and MVAQD. The experimental results demonstrated the efficacy of jointly using local and global qualities. The multi-regression proved to be a bit challenging on OIQA compared to MVAQD, reflecting the necessity to accurately regulate the training process.

Assessing the quality of 360-degree images based on individual regions presents a challenging task. The lack of ground truth opinion scores (MOS) for specific regions makes it difficult to evaluate image quality accurately. Existing datasets only provide MOS for entire 360-degree images, which limits the granularity of assessment. To overcome this challenge, we propose a novel framework that employs adaptive patch labeling techniques. We leverage a set of 2D-IQA methods to generate quality score distributions for each patch in the 360-degree images. These distributions, combined with the available MOS, serve as labels for individual patches, providing a more comprehensive characterization of patch quality. Furthermore, we use these labels to adaptively select and refine deep neural features. By selectively choosing label-specific features, we enhance the accuracy and effectiveness of patch-based 360-degree image quality assessment. This approach allows us to focus on the most relevant and informative features for each patch, resulting in improved assessment performance. The experimental results on two benchmark datasets demonstrate that adaptive patch labeling and feature selection achieve accurate and reliable performances, thus advancing the field of 360-degree image quality assessment.

Currently, there are debates on the accuracy of vision transformers (ViTs) compared to ConvNets for image processing tasks. Image quality assessment (IQA) and particularly 360-IQA is lacking insights regarding their performances and robustness compared to the widely used ConvNets. This paper aims to investigate transfer learning from two pre-trained versions of ViTs and two ConveNets (ResNet-50 and EfficientNet-B3) for 360-degree image quality assessment with a focus on (i) the prediction accuracy and generalization ability and (ii) their adaptation to the specific characteristics of 360-degree images. Furthermore, the influence of adaptive patches sampling compared to simply using equirectangular content is analyzed with each architecture. Experimental findings on publicly available datasets (OIQA, CVIQ and MVAQD) show the superiority of ResNet-50 over ViTs and EfficientNet-B3 while requiring less computational time. Also, the base version of ViTs outperforms the larger one. Finally, except for CVIQ, both ViTs and ConveNets benefit from the adaptive sampling strategy, depicting the interest of taking 360-degree characteristics into account.

Prior to training convolutional neural networks (CNNs) for image quality assessment (IQA), input normalization is sometimes recommended and sometimes not, according to the literature. Although input normalization is known to improve model training and helps in learning important features, it may result in the loss of information such as contrast, color, and luminance. To better explore this issue, we conduct an empirical study to first investigate the effect of normalization on model performance and then which normalization method best fits IQA among existing methods. The performances of the selected methods are statistically compared with three basic scaling methods. The application of normalization is found to be statistically significant on three IQA databases. The performance improvement on the overall databases, as well as per-individual degradation, is demonstrated in the experimental results.

In this paper, we conduct an extensive study on the use of pre-trained convolutional neural networks (CNNs) for omnidirectional image quality assessment (IQA). To cope with the lack of available IQA databases, transfer learning from seven pre-trained CNN models is investigated over retraining on standard 2D databases. In addition, we explore the influence of various image representations and training strategies on the model’s performance. A comparison of the use of projected versus radial content, and multichannel CNN versus patch-wise training is also covered. The experimental results on two publicly available databases are used to draw conclusions about which strategy best fits the visual quality prediction and at which computational cost. The analysis shows that retraining CNN models on 2D IQA databases improves the prediction accuracy. The latter and the required computational time are found to be significantly affected by the training strategy. Cross-database evaluations demonstrate that the nature and variety of the content impact the generalization ability of the models. Finally, we show that conclusions coming from other image processing communities may not hold for IQA. The provided discussion shall provide insights and recommendations when using pre-trained CNNs for omnidirectional IQA.

Image quality assessment of immersive content and more specifically 360-degree one is still in its infancy. There are many challenges regarding sphere vs. projected representation, human visual system (HVS) properties in a 360-degree environment, etc. In this paper, we propose the use of CNNs to design a no reference model to predict visual quality of 360-degree images. Instead of feeding the CNN with ERPs, visually important viewports are extracted based on visual scan-path prediction and given to a multi-channel CNN using DenseNet-121. Moreover, information about visual fixations and just noticeable difference are used to account for the HVS properties and make the network closer to human judgment. The scan-path is also used to create multiple instances of the database so as to perform a robust generalization analysis and compensate for the lack of databases.

The use of convolutional neural networks (CNN) for image quality assessment (IQA) becomes many researcher’s focus. Various pre-trained models are fine-tuned and used for this task. In this paper, we conduct a benchmark study of seven state-of-the-art pre-trained models for IQA of omnidirectional images. To this end, we first train these models using an omnidirectional database and compare their performance with the pre-trained versions. Then, we compare the use of viewports versus equirectangular (ERP) images as inputs to the models. Finally, for the viewports-based models, we explore the impact of the input number of viewports on the models’ performance. Experimental results demonstrated the performance gain of the re-trained CNNs compared to their pre-trained versions. Also, the viewports-based approach outperformed the ERP-based one independently of the number of selected views.