Reference-Free Image Quality Metric for Degradation and Reconstruction Artifacts

TL;DR

The paper tackles reference-free image quality assessment by introducing a self-supervised Quality Factor (QF) Predictor, a seven-layer fully convolutional network that learns to predict the JPEG Quality Factor $QF$ from compressed patches. Trained with a JPEG-based self-supervision objective, the model outputs a $QF$ map and generalizes to non-JPEG artifacts such as Gaussian blur and noise, as well as MRI under-sampling scenarios. Key contributions include a designed training regime with a weighted $QF$ distribution, a regression formulation over classification, and demonstrations of the model as a perceptual loss term in restoration networks, with caveats about balancing with data fidelity. The method provides a lightweight, reference-free proxy for perceptual quality, enabling artifact measurement and dataset quality assessment when ground-truth references are unavailable.

Abstract

Image Quality Assessment (IQA) is essential in various Computer Vision tasks such as image deblurring and super-resolution. However, most IQA methods require reference images, which are not always available. While there are some reference-free IQA metrics, they have limitations in simulating human perception and discerning subtle image quality variations. We hypothesize that the JPEG quality factor is representatives of image quality measurement, and a well-trained neural network can learn to accurately evaluate image quality without requiring a clean reference, as it can recognize image degradation artifacts based on prior knowledge. Thus, we developed a reference-free quality evaluation network, dubbed "Quality Factor (QF) Predictor", which does not require any reference. Our QF Predictor is a lightweight, fully convolutional network comprising seven layers. The model is trained in a self-supervised manner: it receives JPEG compressed image patch with a random QF as input, is trained to accurately predict the corresponding QF. We demonstrate the versatility of the model by applying it to various tasks. First, our QF Predictor can generalize to measure the severity of various image artifacts, such as Gaussian Blur and Gaussian noise. Second, we show that the QF Predictor can be trained to predict the undersampling rate of images reconstructed from Magnetic Resonance Imaging (MRI) data.

Figures (17)

• Figure 1: Examples that illustrate the discrepancy between PSNR values and human perception, adopted fromAuthors01. Given the original image for reference, the left image has a higher metric score but a lower perceptual quality. On the contrary, the middle image has a lower metric score but a higher perceptual quality.
• Figure 2: Showcase of JPEG Compression artifacts within different QFs.Authors09
• Figure 3: Training pipeline for QF Predictor
• Figure 4: QF Predictor Architecture
• Figure 5: Examples of activation maps extracted from the last convolution layer. These results indicate that the network over-emphasizes the grid-like patterns in the corrupted images, which leads to biased results toward low QFs.