IQA

Image Quality Assessment (IQA) is the field that studies how to quantify the perceptual quality of images. It involves methods that assign numerical scores to distorted or processed images to reflect how closely they match a reference image or how they are perceived by human observers. IQA is used across the image processing pipeline, including compression, transmission, restoration, acquisition, and display evaluation.

IQA approaches are commonly categorized by reference availability. Full-reference IQA (FR-IQA) requires a pristine reference image for comparison. Reduced-reference IQA (RR-IQA) uses partial information from the reference, while no-reference or blind IQA (NR-IQA) estimates quality without any reference. Objective IQA scores are often evaluated against human judgments, typically expressed as Mean Opinion Scores (MOS) or differential MOS (DMOS).

Classic metrics include PSNR and SSIM, with SSIM and its multiscale variant (MS-SSIM) offering better alignment with human perception. Other traditional measures include VIF and information-theoretic or gradient-based methods. In recent years, learning-based and deep perceptual metrics such as LPIPS and other neural network–driven scores have become influential, frequently achieving higher correlation with MOS on diverse datasets.

Common IQA benchmarks use public datasets such as LIVE, CSIQ, TID2013, and KADID-10k, which provide images with controlled distortions and human quality judgments. These resources support the development and evaluation of both traditional and learning-based IQA methods.

Applications of IQA include optimizing image and video codecs, assessing transmission and display pipelines, guiding restoration and enhancement algorithms, and informing camera and sensor design.