Large Kernel Distillation Network for Efficient Single Image Super-Resolution

​

Chengxing Xie^1∗ Xiaoming Zhang^1∗ Linze Li¹ Haiteng Meng¹

Tianlin Zhang² Tianrui Li¹ Xiaole Zhao^1†

�¹ Southwest Jiaotong University, China�² National Space Science Center, Chinese Academy of Science, China

Super-Resolution

downsample

upsample

GT

LR

SR

ill-posed

Motivation

[1] Guo, M. H., Lu, C. Z., Liu, Z. N., Cheng, M. M., & Hu, S. M. (2022). Visual attention network. arXiv preprint arXiv:2202.09741.

Large-Kernel Decomposition

Motivation

[2] Ding, X., Zhang, X., Han, J., & Ding, G. (2021). Diverse branch block: Building a convolution as an inception-like unit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 10886-10895).886-10895.

Structural Re-parameterization

Architecture

Shallow Feature Extraction

Deep Feature Extraction

Image Reconstruction

Architecture

Down-Sample

Up-Sample

Conv

Sigmoid

Sigmoid

Conv Groups

Contrast

Sigmoid

Conv Groups

Enhanced spatial attention (ESA)

Contrast-aware channel attention (CCA)

[3] Hui, Z., Gao, X., Yang, Y., & Wang, X. (2019, October). Lightweight image super-resolution with information multi-distillation network. In Proceedings of the 27th acm international conference on multimedia (pp. 2024-2032).

[4] Liu, J., Zhang, W., Tang, Y., Tang, J., & Wu, G. (2020). Residual feature aggregation network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 2359-2368). 2020: 2359-2368.

Architecture

(a) LKDB: Large Kernel Distillation Block; (b) BSConv: Blueprint Separable Convolution; (c) LKA: Large Kernel Attention; (d) RBSB: Re-parameterized Blueprint Shallow Block.

Optimizer

Heavy-ball acceleration (HBA)

Nesterov Accelerated Gradient (NAG)

• To compute the “ahead-of-time gradient.”, Adan can pre-perceive the geometric information around the current point.

​

• Adan can adapt to larger learning rates and batch sizes, leading to faster convergence.

Adam

Adan

[5] Xie, X., Zhou, P., Li, H., Lin, Z., & Yan, S. (2022). Adan: Adaptive Nesterov Momentum Algorithm for Faster Optimizing Deep Models. arXiv preprint arXiv:2208.06677.

Ablation Studies

• Removing ESA&CCA can lead to a large decrease in performance.

​

• Adding LKA can significantly improve performance with a small number of additional parameters.

​

• Adjusting the number of network channels can slightly increase performance.

Table 1. Ablation study on large kernel attention.

Ablation Studies

Table 2. PSNR / SSIM comparison of different basic blocks in the feature distillation connections of LKDN-S.

• Removing shortcut and using structural re-parameterization can increase performance.

Table 3. PSNR / SSIM comparison of applying Adam and Adan optimizers.

Ablation Studies

• Adan optimizer has faster rate of convergence and shorter training time.

​

• Adam optimizer is less likely to fall into local optima than the Adam optimizer.

Performance

Table 5. Comparison with state-of-the-art methods, and our training dataset is DF2K (2650 images).

Performance

Comparison of model performance and complexity on Urban100 with SR(×4).

Performance

• Clear lines

​

• Less artifacts

Thank you for your attention.

Q&A