The aesthetic and technical quality of images is an important factor in providing users with a better visual experience. Image Quality Assessment (IQA) allows models to build a relationship between an image and the user’s perception of its quality. Many IQA approaches have achieved success using the idea of convolutional neural networks (CNNs). In addition, IQA models based on CHN are limited by the fixed-size input requirements of batch learning, that is, input images that need to be cropped or resized to a fixed-size shape.
To address this problem, researchers at Google presented the ‘Multiscale Image Quality Transformer (MUSIQ)’ to circumvent CNN’s fixed-size input limitations to predict effective image quality on native-resolution images. The paper was published at ICCV 2021, where the model supports processing full-size input images at various resolutions and aspect ratios. It will also allow multiscale feature extraction to produce image quality with different granularities.
“We are applying MUSIQ to four large-scale IQA datasets, demonstrating consistent state-of-the-art results across three technical quality datasets (PaQ-2-PiQ, KonIQ-10k and SPAQ) and comparable to the performance of current models across the AVA aesthetic quality dataset,” the study says.