Abstract
Blind motion deblurring remains a challenging ill-posed problem in computer vision, often requiring complex multi-scale architectures to restore sharp images from unknown blur kernels. Traditional multi-stage models frequently suffer from high computational overhead and the loss of high-frequency details during downsampling. This paper presents MLWNet, an efficient multi-scale network that leverages a Learnable Discrete Wavelet Transform (LWN) and a Single-Input Multiple-Output (SIMO) architecture. By replacing standard fusion modules with learnable wavelet-based feature extraction, MLWNet achieves state-of-the-art performance on real-world datasets while significantly reducing complexity. Our results demonstrate that MLWNet is particularly effective at capturing the directional continuity of real-world motion blur, outperforming existing transformers and multi-stage models in both accuracy and efficiency.
1. Introduction
Blind motion deblurring is the process of recovering a sharp latent image from a single blurry observation where the blur kernel is unknown. To handle the varying scales of blur, modern deep learning approaches typically adopt a coarse-to-fine strategy. However, these Multiple-Input Multiple-Output (MIMO) architectures often rely on manually downsampled image pairs and computationally expensive fusion modules to bridge the gap between pixel-level information and deep semantic features. Furthermore, standard downsampling operations often discard the high-frequency information—such as edges and textures—that is critical for successful restoration.
To address these limitations, we introduce MLWNet. Our approach simplifies the multi-scale pipeline by utilizing a SIMO architecture and a novel learnable wavelet transform that preserves spatial localization and directional frequency data, ensuring that critical edge information is maintained throughout the restoration process.
2. Proposed Method: MLWNet
2.1 SIMO Architecture
Unlike traditional multi-scale models that require a stack of pre-processed, downsampled images as input, MLWNet adopts a Single-Input Multiple-Output (SIMO) design. The network takes a single full-resolution image and extracts features at various scales internally. During training, it generates sharp outputs at multiple resolutions, which supervises the learning process across scales without the need for redundant manual preprocessing. This architecture significantly reduces the computational burden compared to multi-stage or MIMO models.
2.2 Learnable Discrete Wavelet Transform (LWN)
The core innovation of MLWNet is the Learnable Discrete Wavelet Transform (LWN). While Fourier Transforms have gained popularity for frequency-domain analysis in restoration, they often lack the spatial localization necessary for non-uniform blur. Wavelets, conversely, provide excellent spatial and frequency localization, making them ideal for identifying abrupt signals like sharp edges.
The LWN module is implemented using group convolutions where the filters are learnable but constrained by a self-supervised Wavelet Loss. This allows the network to adaptively decompose features into four sub-bands (Low-Low, Low-High, High-Low, and High-High). By doing so, the model explicitly learns to recognize and compensate for blur trajectories along horizontal, vertical, and diagonal axes.
2.3 Network Components
The architecture is built upon three primary blocks:
- Simple Encoder Blocks (SEB): Responsible for efficient feature extraction.
- Wavelet Fusion Blocks (WFB): Facilitate the merging of semantic information across different scales.
- Wavelet Head Blocks (WHB): Perform the final image restoration at each respective scale.
3. Experimental Evaluation
3.1 Datasets and Metrics
We evaluated MLWNet on several benchmark datasets representing both real-world and synthetic blur:
- RealBlur-J & RealBlur-R: Large-scale datasets of real-world blurry/sharp pairs.
- RSBlur: A dataset focused on realistic blur synthesis.
- GoPro: A standard synthetic dataset generated via frame averaging.
3.2 Performance and Efficiency
MLWNet sets a new benchmark for efficiency and restoration quality. On the RealBlur-J dataset, the MLWNet-B variant achieved a PSNR of 33.84 dB, representing a 1.02 dB improvement over the previous state-of-the-art (GRL).
Beyond accuracy, MLWNet is remarkably efficient. It achieves higher PSNR with significantly fewer Giga-Multiply-Accumulate operations (GMACs) than competing multi-stage transformers. For instance, the model maintains a runtime of approximately 0.05s for a 256x256 patch, making it suitable for practical applications where computational resources are limited.
4. Discussion: The Synthetic Gap
An interesting finding in our research is that MLWNet performs better on real-world blur than on synthetic datasets like GoPro. This "synthetic gap" arises because synthetic blur often contains discontinuous or "steppy" trajectories due to the discrete nature of frame averaging. Because MLWNet’s wavelet-based approach is optimized for the continuous physical motion found in the real world, it is uniquely robust to natural blur but can struggle with the unnatural artifacts present in some synthetic benchmarks.
5. Conclusion
MLWNet demonstrates that a well-designed wavelet-based SIMO architecture can outperform more complex MIMO models in blind motion deblurring. By making the Discrete Wavelet Transform learnable and integrating it into a multi-scale framework, we provide a solution that is both more accurate for real-world motion and significantly more efficient. Future work will explore extending this architecture to other restoration tasks, such as low-light enhancement and super-resolution, where frequency-aware feature extraction is equally vital.
Summary updated