Role of Visual Saliency in Video Quality Assessments

The human visual system naturally prioritizes unique and salient objects within a scene. In computer vision, visual saliency refers to the property that makes specific regions stand out in an image or video. With the growing reliance on video-based applications such as streaming, virtual reality, and video conferencing, video quality assessment (VQA) has become crucial. This research investigates the role of Global Contrast-Based Visual Saliency in VQA and proposes two methodologies: 1) feature extraction using Visual Saliency, 2) Visual Saliency combined with Feature Extraction and Feature Selection. The methodologies were evaluated using the LIVE Video Quality Challenge Database and multiple regression models. Experimental results demonstrate that incorporating visual saliency significantly enhances VQA accuracy while reducing computational complexity. The best-performing approach—Visual Saliency combined with Feature Extraction and Feature Selection—achieved the highest correlation with subjective quality scores (PCC = 0.997, KRCC = 0.955, SROCC = 0.995) and the lowest Mean Squared Error (MSE = 1.318). This highlights the effectiveness of visual saliency and feature selection in improving the reliability and efficiency of VQA models, making them more aligned with human perception.