Focused Frames: Enhancing Video Quality Assessment Through Perceptual Pre-Cropping

Automatic Video Quality Assessment (VQA) models often struggle to match human perception because they evaluate entire frames uniformly, neglecting the impact of visual attention on specific Regions of Interest (ROI). To bridge this gap, we propose a novel, training-free method that enhances existing VQA architectures by aligning them with perceptual focus. Leveraging a deep learning video cropper pre-trained on human-annotated saliency data, we extract practical bounding boxes that represent real-world ROIs. We introduce a lightweight fusion strategy that combines quality information from both the global context and the perceptually focused cropped view, without requiring model retraining. Experimental validation on standard datasets, including LIVE-VQC and KoNViD- 1k, demonstrates that this approach yields consistent improvements in Spearman’s Rank-Order Correlation Coefficient (SROCC) and Pearson’s Linear Correlation Coefficient (PLCC) across multiple no-reference VQA models. These results confirm that our plug-and-play method effectively increases predictive accuracy and can be seamlessly integrated into contemporary video delivery pipelines.