RLKA-Net

Official codes for Paper "Recurrent Large Kernel Attention Network for Efficient Single Infrared Image Super-Resolution"

Environment

• Pytorch >= 1.8
• BasicSR >= 1.3.5

Introduction

Architecture

Overview of the architecture of RLKA-Net

Recurrent Learning Stragety

Experiment Result on CVC09 Dataset (x4)

Training

Dataset:

Original Infrared Images

• OSU
• FLIR
• CVC09
• LLVIP

Train/Test Dataset used in our work

• OSU
• FLIR
• CVC09
• LLVIP

Downsample

Apply bicubic or lanczos downsample to obtain LR infrared images (Scale = 0.25 or 0.5).

' python scripts\infrared_multiscale.py --input \GT_IMAGE_PATH --output \LR_IMAGE_PATH --scale 0.5(or 0.25) --method bicubic(or lanczos)'

Traning command

BasicSR framework is utilized to train our RLKA-Net.

' python basicsr/train.py -opt options/train/RLKAN/train_rlkan_flir_x4_r.yml '

Before running this training command, you should prepared the paired FLIR infrared images.

Pre-trained models

Part of the Pre-trained models is avaliable here (Google Drive)

Infrared SR Performance

Acknowledgements

We'd like to thank MAN and BasicSR for their enlightening work, and thank the author of OSU, FLIR, CVC09 and LLVIP for provide open-source infrared images!