ENAF: A Multi-Exit Network with an Adaptive Patch Fusion for Large Image Super Resolution

November 6, 2024 ยท View on GitHub

This is the sample source code provided for WACV 2025 submission: "ENAF: A Multi-Exit Network with an Adaptive Patch Fusion for Large Image Super Resolution".

enaf

Data preparation

  • The data are prepared following ClassSR
  • Raw data (DIV2K) can be downloaded from DIV2K
  • Testing data Test2K, Test4K, Test8K can be downloaded from Drive

Pretrained weights:

Configuration

  • Adjust configuration and weight path in folder template
  • Importantly, notice the test data dir path.

Inference

  • Run the script placed in each network's script:
$ script/FSRCNN_x4/TEST/1est/fsrcnn_eunaf_test8k.sh  # FSRCNN
$ script/CARN_x4/TEST/1est/carn_eunaf_test8k.sh  # CARN
$ script/SRResNet_x4/TEST/1est/srresnet_eunaf_test8k.sh  # SRResNet

Results

  1. Main result on SISR performance and cost reduction. main_result

  2. Ablation study on effectiveness of PSNR estimator as early exit signal. ablation

Acknowledgement

  • Our implementation based on ClassSR. We thank authors for their valuable contributions.

Contact