BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator, IEEE TIP 2012) is a classic handcrafted no-reference image-quality model built on natural scene statistics, the finding that undistorted natural images obey consistent statistical regularities that blur, noise, and compression measurably break. It works in the spatial domain from locally normalized luminance coefficients, scoring how far a frame has drifted from natural-looking statistics. Unlike the training-free NIQE, BRISQUE is opinion-aware: it is trained on a set of distorted images that humans have rated, so it learns the mapping from statistical deviation to perceived quality. It scores a single frame, so for video it is run per frame and the results are pooled, and it is available through toolboxes such as pyiqa and scikit-video. Its catch is that, tuned on photographic distortions, it correlates poorly with human judgment on in-the-wild user-generated video, with reported cross-dataset SROCC around 0.31 to 0.58, so its number must be validated on content like yours.
