Towards Region-Level No-Reference Image Quality Assessment

Region-level No-Reference Image Quality Assessment (NR-IQA) enables fine-grained quality assessment for user-specified regions, which is critical in applications, such as camera imaging, autonomous driving and image compression. While existing NR-IQA methods have achieved remarkable success in assessing full-image quality, few explore quality analysis for user-specified Regions-of-Interest (ROIs). This paper proposes a novel framework and two datasets for ROI-level NR-IQA. First, we present Seagull, a network that SEes and Assesses ROI quality with GUidance from a Large Language model. Seagull achieves accurate ROI-level quality prediction through three key components: 1) a mask-based ROI for precise region specification, 2) a carefully designed Mask-based Feature Extractor (MFE) that jointly models global contexts and local details, and 3) a powerful large-language model (LLM) to enhance semantic understanding and quality reasoning. Furthermore, to support training and evaluation, we construct two ROI-level IQA datasets: Seagull-1M and Seagull-3K. Seagull-1M contains 1 million synthetically distorted images with 33 million ROIs, designed to strengthen the model's perception of local distortions. Seagull-3K consists of 3,261 real-world distorted ROIs, curated to improve generalization under authentic imaging conditions. Extensive experiments show that Seagull, after pre-training on Seagull-1M and fine-tuning on Seagull-3K, achieves outstanding performance in ROI-level quality assessment, outperforming baseline methods and demonstrating strong generalization across diverse distortion types and domains. Visualization results further indicate that Seagull is able to generalize to in-the-wild distortions and images.