FCL-GAN: A Lightweight and Real-Time Baseline for Unsupervised Blind Image Deblurring

Suiyi Zhao, Zhao Zhang, Richang Hong, Mingliang Xu, Yi Yang, Meng Wang

Introduction

Blind Image Deblurring (BID) aims to recover sharp images from blurred inputs without prior knowledge of the blur kernel. While supervised deep learning models have achieved significant success, they rely heavily on synthetic paired datasets, often failing to generalize to complex real-world "natural" blur. Unsupervised methods, such as those based on CycleGAN, offer a more flexible alternative by learning from unpaired data. However, existing unsupervised models are typically characterized by massive parameter counts and slow inference speeds, hindering their deployment in real-time applications like mobile video processing.

FCL-GAN addresses these challenges by providing a lightweight, real-time framework that is approximately 25x smaller and 5x faster than previous state-of-the-art (SOTA) unsupervised methods.

Methodology

The FCL-GAN architecture introduces two primary innovations to achieve high efficiency without sacrificing restoration quality: the Lightweight Domain Conversion Unit (LDCU) and the Parameter-Free Frequency-Domain Contrastive Unit (PFCU).

Lightweight Domain Conversion Unit (LDCU)

To reduce redundancy found in standard generators, the authors developed a Lightweight Encoder-Decoder (LED) structure. Unlike traditional multi-scale pipelines that are computationally expensive, the LED stacks "residual metas" at higher resolutions to preserve fine spatial details while maintaining a low parameter count. Furthermore, the model utilizes Batch Normalization (BN) instead of the standard Instance Normalization (IN), as BN was found to more effectively capture global domain transitions from blurred to sharp states.

Parameter-Free Frequency-Domain Contrastive Unit (PFCU)

Contrastive learning typically requires additional neural network branches to map images into a latent space, increasing model weight. FCL-GAN bypasses this by operating entirely in the frequency domain using Fast Fourier Transforms (FFT). The PFCU leverages the observation that blurred images lose high-frequency signals compared to sharp images. By implementing a "black coverage" similarity measure in the frequency domain, the model enforces sharpness constraints without adding any trainable parameters.

The FCL-GAN architecture featuring the Lightweight Domain Conversion Unit (LDCU) and the Parameter-Free Frequency-Domain Contrastive Unit (PFCU).

Experiments and Results

FCL-GAN was evaluated across several benchmarks, including GoPro, HIDE, CelebA, and RealBlur. The results demonstrate that FCL-GAN sets a new standard for efficiency in unsupervised deblurring.

Quantitative Performance

As shown in the performance comparison, FCL-GAN achieves superior metrics across PSNR, SSIM, and NIQE while drastically reducing resource requirements. Specifically, the model size is reduced to 24.56 MB (vs. 606.8 MB for SOTA), and runtime for a 720p image is cut to 10.86 ms (vs. 62.06 ms for SOTA). This runtime is well below the 33.33 ms threshold required for 30 FPS real-time video processing.

Comparison of FCL-GAN (Ours) against SOTA in terms of performance metrics, model size, and runtime.

Key Insights

  • Frequency Domain Constraints: Utilizing frequency-domain information provides a computationally "free" yet highly effective way to guide the deblurring process.
  • Normalization Strategy: Switching to Batch Normalization proved crucial for the model's ability to learn the global "blur-to-sharp" transition effectively in an unsupervised setting.

Conclusion

FCL-GAN demonstrates that high-quality unsupervised deblurring does not require massive computational overhead. By optimizing the generator architecture and introducing parameter-free frequency-domain constraints, the authors have created a robust baseline for real-time applications. While a gap remains between unsupervised and the best fully supervised models, FCL-GAN's efficiency and generalization capabilities make it a highly practical solution for real-world scenarios. Future work will explore further hardware-specific optimizations and advanced frequency-domain strategies.

Summary updated

Abstract

Blind image deblurring (BID) remains a challenging and significant task. Benefiting from the strong fitting ability of deep learning, paired data-driven supervised BID methods have obtained great progress. However, paired data are usually synthesized by hand, and the realistic blurs are more complex than synthetic ones, which makes the supervised methods inept at modeling realistic blurs and hinders real-world applications. As such, unsupervised deep BID methods without paired data offer certain advantages, but current methods still suffer from some drawbacks, e.g., bulky model size, long inference time, and strict image resolution and domain requirements. In this paper, we propose a lightweight and real-time unsupervised BID baseline, termed Frequency-domain Contrastive Loss Constrained Lightweight CycleGAN (shortly, FCL-GAN), with attractive properties, i.e., no image domain limitation, no image resolution limitation, 25x lighter than SOTA, and 5x faster than SOTA. To guarantee the lightweight property and performance superiority, two new collaboration units called lightweight domain conversion unit (LDCU) and parameter-free frequency-domain contrastive unit (PFCU) are designed. LDCU mainly implements inter-domain conversion in lightweight manner. PFCU further explores the similarity measure, external difference and internal connection between the blurred domain and sharp domain images in frequency domain, without involving extra parameters. Extensive experiments on several image datasets demonstrate the effectiveness of our FCL-GAN in terms of performance, model size and inference time.
BibTeX
@inproceedings{zhao2022fcl,
  author = {Suiyi Zhao and Zhao Zhang and Richang Hong and Mingliang Xu and Yi Yang and Meng Wang},
  title = {{FCL-GAN: A Lightweight and Real-Time Baseline for Unsupervised Blind Image Deblurring}},
  booktitle = {ACM Multimedia},
  year = {2022},
  url = {https://openreview.net/forum?id=r8tbK9Ugg4},
}
Figure