Judging photo quality without shrinking the image

Evaluating image quality on massive 4K+ photos usually requires shrinking them (losing fine details) or sampling isolated crops (losing context). This work treats image quality assessment as a graph problem: sample patches from the full image, connect them based on spatial proximity and feature similarity, then use graph convolution to let each patch learn from its neighbors. On the UHD-IQA benchmark, it achieves RMSE of 0.0519—better absolute accuracy than competing methods—while preserving both fine detail and scene context.