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

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".

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:

• Weights are stored in 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.

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

Acknowledgement

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

Contact

• Email: duong.nm210243@sis.hust.edu.vn