Abstract
Motion deblurring remains a challenging ill-posed problem due to the loss of fine-grained details and training instabilities in deep generative models. This paper introduces Bi-Skip, a novel framework that combines a specialized dual-path generator architecture with Self-paced Learning (SPL). By extracting both shallow and deep image priors and employing a curriculum-based training strategy, Bi-Skip achieves state-of-the-art restoration quality and significantly faster inference speeds compared to existing methods.
Introduction
Restoring a sharp image from a single blurry observation is a classic challenge in computer vision. Traditional methods often rely on brittle blur kernel estimation, leading to ringing artifacts. While modern Deep Learning approaches have improved results, they frequently struggle with feature loss during deep extraction and training instability when faced with complex motion blur. Bi-Skip addresses these issues through architectural innovation and a robust training regime.
Methods
1. The Bi-Skip Architecture
Standard architectures like U-Nets often lose shallow structural details when bottlenecking features. The Bi-Skip network mitigates this by extracting and concatenating both shallow features (via $1 \times 1$ convolutions) and deep features (via Residual Blocks) at every scale. This dual-path approach allows the decoder to simultaneously process structural priors and fine details.
2. Bi-level Loss Function
To balance mathematical accuracy with visual appeal, the authors employ a bi-level loss function. It explicitly combines a normalized $L_1$ pixel loss for fidelity with a VGG19-based perceptual loss to ensure the output remains visually pleasing and free of common artifacts.
3. Self-paced Learning (SPL)
The core of the training strategy is a "curriculum" approach. Instead of training on all samples simultaneously, the SPL mechanism starts with "easy" samples (low loss). As the network matures, an "age" parameter gradually introduces more complex, "harder" blur samples. This prevents early training perturbations and leads to a more robust model capable of handling extreme motion.
Experimental Results
Performance and Efficiency
Evaluated on the GoPro and Köhler datasets, Bi-Skip outperformed several state-of-the-art methods. On the GoPro benchmark, it achieved a PSNR of 30.37 dB and an SSIM of 0.9541. Crucially, the model is highly efficient, processing a single image in just 0.76 seconds—more than twice as fast as many competing high-performance models.
Saliency and Ablation
Ablation studies confirmed that the SPL mechanism (SA1P-BS) provides a substantial performance boost over the base model. Furthermore, eye-tracking data indicated that the model preserves details in "saliency regions" where human attention is naturally focused, enhancing the perceived quality of the restoration.
Conclusion
Bi-Skip demonstrates that combining specialized architectural priors with a self-paced training curriculum can overcome the traditional trade-offs between speed and quality in motion deblurring. While challenges remain in bridging the gap between synthetic and real-world blur, the framework provides a robust foundation for real-time, high-fidelity image restoration.
Summary updated