QuARF: Quality-Adaptive Receptive Fields for Degraded Image Perception

Fei Gao, Ying Zhou, Ziyun Li, Wenwang Han, Jiaqi Shi, Maoying Qiao, Jinlan Xu, Nannan WangAAAI 2025 · DOI

TL;DR

This paper introduces QUARF, a novel plug-and-play module that dynamically adjusts a neural network's receptive fields based on the quality of the input image. By doing so, QUARF significantly improves performance on degraded images across various computer vision tasks, outperforming traditional data augmentation methods.

What Problem Does This Paper Address?

Deep neural networks (DNNs) excel at processing high-quality images but suffer a significant performance drop when faced with degraded inputs (e.g., blurry, noisy). While data augmentation is a common solution, it can degrade performance on pristine images. This paper addresses the critical need for DNNs that are robust to varying image quality without sacrificing performance on clean data, a challenge that hinders their real-world applicability. The authors are motivated by the human visual system's ability to adapt its perception based on image quality, suggesting that artificial systems should do the same.

Methods

The core innovation is the Quality-Adaptive Receptive Fields (QUARF) module. It's designed to be easily integrated into existing DNN architectures. QUARF consists of two main components:

  1. Multi-Kernel Convolutional Block (MRF): This block employs multiple convolutional kernels with varying sizes (from 3x3 up to 17x17 in their experiments), creating a range of receptive fields. To manage computational cost, they use dilated convolutions. The kernels are designed to be "continuous" in the frequency domain, covering a broad spectrum of visual information.
  2. Quality-Adaptive Routing (QAR): This is the "smart" part. For each input image, a Blind Image Quality Assessment (BIQA) model (they default to CONTRIQUE) extracts quality features. These features are then fed into a shallow Multi-Layer Perceptron (MLP) to predict the "significance" or weight for each receptive field within the MRF. This allows the network to dynamically select the most appropriate receptive fields for the given image quality. To enhance efficiency, the input feature map is split into multiple groups, with each group learning its own quality-adaptive routing parameters.

Datasets

The paper evaluates QUARF on a variety of established datasets across different tasks:

  • Image-to-Image Translation (I2IT) & Neural Style Transfer (NST): CUFS, summer2winter, Metface, COCO, WikiArt.
  • Image Restoration: AGAN-Data, BSD100, Manga109, Urban100, DF1K.
  • Face Parsing: CelebAMask-HQ.
  • Semantic Segmentation: HQSeg-44K, ThinObject-5K, DIS, HR-SOD, COIFT.

They do not introduce new datasets but rather demonstrate QUARF's effectiveness on existing benchmarks.

Key Insights and Ideas

  • Receptive Field Adaptability is Key: The central idea is that optimal feature extraction depends on image quality. Degraded images require different receptive field configurations than pristine ones.
  • Mimicking Human Perception: The approach is inspired by how the human visual system uses multiscale processing and adapts to image clarity.
  • Beyond Data Augmentation: QUARF offers a more principled way to handle degraded inputs, avoiding the performance trade-offs associated with aggressive data augmentation.
  • Plug-and-Play Simplicity: The module's design allows it to be easily integrated into various existing DNN architectures without major modifications.
  • Quality Features Drive Selection: The effectiveness of the quality-adaptive routing hinges on accurately extracting image quality information.

Key Results

QUARF consistently demonstrates significant improvements across diverse tasks and baselines:

  • Image Translation: Achieved the lowest FID (Fréchet Inception Score) and competitive SIFID (Single Image Fréchet Inception Distance) and CPBD (Cumulative Probability of Blur Detection) scores, indicating better style realism and clarity. For example, on MATEBIT, QUARF achieved an FID of 29.87, significantly lower than Data Aug. (83.97).
  • Image Restoration: Showed stable and significant boosts in PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) for super-resolution tasks, and improved visual quality with fewer artifacts in deraining.
  • Face Parsing: Led to preciser parsing results and better performance on degraded images, particularly in capturing regional boundaries, as evidenced by improved mIoU (mean Intersection over Union) and BIoU (Boundary IoU).
  • Semantic Segmentation: Reduced prediction errors on low-quality images and achieved optimal BIoU values, resulting in more continuous segmentation regions with precise boundaries.

Limitations and Future Work

  • BIQA Model Choice: While QUARF is flexible regarding the BIQA model used, the performance can be influenced by the quality of the extracted features. The paper suggests that using better quality representations could further boost performance.
  • Computational Overhead: Although they use dilated convolutions and group splitting for efficiency, QUARF does introduce some computational overhead compared to a standard convolutional layer.
  • Future Work: The authors plan to explore more advanced approaches for processing, analyzing, and generating tasks involving degraded images, potentially by extending QUARF's capabilities. They also highlight the potential for further improvements by exploring different BIQA methods and optimizing the number of groups and kernel sizes.

Summary updated

BibTeX
@inproceedings{gao2025quarf,
  author = {Fei Gao and Ying Zhou and Ziyun Li and Wenwang Han and Jiaqi Shi and Maoying Qiao and Jinlan Xu and Nannan Wang},
  title = {{QuARF: Quality-Adaptive Receptive Fields for Degraded Image Perception}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i16.33836},
}
Figure