← Back
IEEE ECICE 2023 Computer Vision ↗ Paper

TMSR: Tiny Multi-path CNNs for Super Resolution

Chia-Hung Liu, Tzu-Hsin Hsieh, Kuan-Yu Huang, Pei-Yin Chen

TMSR CNN architecture

Abstract

TMSR (Tiny Multi-path Super Resolution) proposes a lightweight convolutional neural network architecture for single-image super resolution, optimized for hardware-constrained edge devices. The multi-path design extracts features at multiple receptive field scales simultaneously, fusing them through a channel-attention aggregation module to reconstruct high-frequency image detail efficiently.

Experiments on benchmark SR datasets demonstrate competitive PSNR/SSIM scores with significantly fewer parameters than baseline models, making TMSR suitable for deployment on embedded systems and mobile hardware.

Key Contributions

  • Multi-path convolutional feature extraction with parallel branches at 3×3, 5×5, and dilated kernels.
  • Channel-attention module for adaptive feature fusion across paths.
  • Pixel-shuffle upsampling for sub-pixel reconstruction without checkerboard artifacts.
  • Achieves competitive PSNR on DIV2K, Set5, and Set14 at <200K parameters.

Technologies

PyTorchCNNSuper Resolution Channel AttentionEdge AI Computer VisionIEEE ECICE 2023

BibTeX

@inproceedings{liu2023tmsr,
  title={TMSR: Tiny Multi-path CNNs for Super Resolution},
  author={Liu, Chia-Hung and Hsieh, Tzu-Hsin and Huang, Kuan-Yu and Chen, Pei-Yin},
  booktitle={2023 IEEE 5th Eurasia Conference on IOT, Communication and Engineering (ECICE)},
  pages={829--833},
  year={2023},
  organization={IEEE}
}