A Unified Hybrid Framework for Full-Reference and No-Reference Quality Assessment

Practical deployment of Image Quality Assessment (IQA) systems increasingly demands models that handle both full-reference (FR) and no-reference (NR) scenarios within a single deployable architecture, without compromising accuracy in either mode. Existing frameworks such as DisQUE deliver strong FR performance but are fundamentally constrained to settings where a pristine reference image accompanies every query. We address this gap by introducing a hybrid IQA framework that extends DisQUE into a dual-mode system through a novel per-sample hybrid routing mechanism. At inference time, a binary mask derived from the reference tensor automatically directs each sample to the FR fusion head when a reference is available, or to the NR fusion head otherwise, processing mixed FR/NR batches in a single forward pass without dataset-specific branching. The architecture consists of three shared encoders (CONTRIQUE, ReIQA quality-aware, ReIQA content-aware) whose outputs feed into mode-specific FR and NR fusion heads, followed by a shared regression head. This design separates the mode-specific integration layers from the shared feature extraction backbone, offering practical advantages over deploying two independent models: the encoders are loaded once, memory is shared across modes, and switching between FR and NR inference requires only routing the reference tensor, not reloading weights. The NR pathway benefits from cross-modal knowledge distillation from the frozen DisQUE teacher, governed by \(\mathcal {L}_{\text {total}} = \mathcal {L}_{\text {MOS}} + \lambda _{\text {feat}}\mathcal {L}_{\text {feat}} + \lambda _{\text {score}}\mathcal {L}_{\text {score}}\) with \(\lambda _{\text {feat}} = \lambda _{\text {score}} = 0.5\) . Critically, distillation losses are computed only for FR samples with real reference images; NR samples are supervised exclusively by \(\mathcal {L}_{\text {MOS}}\) . In FR mode, the framework achieves SROCC=0.9470/PLCC=0.9460/RMSE=0.0447 on TID2013 and SROCC=0.9534/PLCC=0.9577/RMSE=0.0698 on KADID-10k, attaining state-of-the-art FR accuracy among all compared methods. In NR mode, the model achieves SROCC=0.8842/PLCC=0.9098/RMSE=0.0499 on KonIQ-10k and SROCC=0.8531/PLCC=0.8735/RMSE=0.1242 on LIVE-Challenge, performing competitively with leading NR methods. Comprehensive ablation studies, cross-dataset generalisation experiments, statistical significance tests, and runtime benchmarks validate each design decision.