The resulting pipeline lets well-established full-reference measures operate in genuinely blind settings, and we analyse when the restored reference is faithful enough for the borrowed metric to stay reliable.

ARIQA-3DS is the first large stereoscopic AR image quality assessment dataset. Its 1,200 AR viewports fuse high-resolution stereoscopic omnidirectional captures of real-world scenes with diverse augmented foregrounds under controlled transparency and degradation conditions. A subjective study with 36 participants using a video see-through head-mounted display collected both quality ratings and simulator-sickness indicators.

Our analysis shows that perceived quality is primarily driven by foreground degradations and modulated by transparency levels, while oculomotor and disorientation symptoms increase progressively but manageably during viewing. The dataset is released publicly as a benchmark for next-generation AR quality models.

This paper presents a deep-learning objective metric designed specifically for AR image quality. The approach has four steps: (1) fine-tuning a self-supervised pre-trained vision transformer to extract prominent features from reference images and distilling that knowledge to improve representations of distorted images, (2) quantifying distortions by computing shift representations, (3) employing cross-attention-based decoders to capture perceptual quality features, and (4) integrating regularisation and label smoothing to address overfitting.

Extensive experiments on the ARIQA dataset show that all model variants — TransformAR, TransformAR-KD and TransformAR-KD+ — outperform existing state-of-the-art methods.

The architecture keeps the transformer’s ability to relate foreground and background regions — the source of visual confusion in AR — but trims capacity aggressively, and is validated against subjective AR quality datasets.

Rather than treating the viewport as an ordinary photograph, the model is designed around the distortions that arise in immersive pipelines — transparency, blending artefacts and the perceptual interference between real and virtual layers — and predicts quality without access to any reference.

We introduce a ranking-guided training strategy for lightweight quality prediction: perceptual ordering feedback is folded into the objective so that the model is rewarded for getting pairwise comparisons right, not just for minimising score error. The result is better calibration and stronger pairwise consistency at a small model size.

This is the conference paper that TransformAR extends: it establishes that a compact transformer, trained with the right supervision, is sufficient to model the interaction between the augmented foreground and the captured background.

We keep the Swin Transformer’s capacity to capture long-range spatial dependencies while refining the representation of clinically salient regions. A localisation-aware refinement module aligns hierarchical transformer tokens with the final classifier distribution, so local cues stay discriminative across every scale. Validation on two public benchmarks shows improved robustness and precision for Kellgren–Lawrence severity prediction.

This chapter surveys the state of the art in copy-move detection across feature representations, learning strategies and public datasets. We catalogue low-level descriptors, machine-learning pipelines and deep architectures, and discuss how each responds to common post-processing such as translation, rotation and scaling. We close on the open challenges in building explainable, trustworthy forensic systems.

We propose an automated framework built on a Swin Transformer backbone with a multi-head prediction module, capturing both global context and fine-grained biomarkers in radiographic images. The model trains on publicly available datasets labelled with Kellgren–Lawrence scores and incorporates a domain adaptation strategy to mitigate dataset shift. Experiments show reliable severity predictions that generalise across data sources.