MDA-SR

This repository is an official PyTorch implementation of the paper "MDA-SR: Multi-level Domain Adaptation Super-Resolution for Wireless Capsule Endoscopy Images"

🔥: Dependencies

• Python 3.7
• PyTorch >= 1.7.0
• matplotlib
• yaml
• importlib
• functools
• scipy
• numpy
• tqdm
• PIL

In this project, we propose a multi-level domain adaptation training framework for the SR of capsule endoscopy images.

🚉: Pre-Trained Models

To achieve SR of capsule endoscopy images, download these 2x, 4x models, and place them in the path ./experiment/ckpt

🚋: Training

We first train adaptive downsampling model alone for 50 epochs, and then train domain adaptation SR model together for 50 epoch. The detailed training command as here:

CUDA_VISIBLE_DEVICE=0 python train.py --name {EXP_PATH} --scale {SCALE} --adv_w 0.01 --batch_size 10 --patch_size_down 256 --decay_batch_size_sr 400000 --decay_batch_size_down 50000 --epochs_sr_start 51 --gpu cuda:0 --sr_model endosr --training_type endosr --joint --save_results --save_log

with following options:

• EXP_PATH is the folder name of experiment results
• scale is the scale of the SR
• adv_w is th hyperparameter. (default: `0.01)

🧩: Evaluation

The detailed evaluation command as here: test sr: generation of super-resolution images

CUDA_VISIBLE_DEVICE=0 python predict.py --test_mode sr_patch --name scale_4x --scale 4 --crop 480 --pretrain_sr ./experiment/ckpt/scale_4x/model_sr_0110.pth --test_lr Capsule_Data/TestSet/side_480 --gpu cuda:6 --sr_model endosr --training_type endosr --save_results --realsr

test down: generation of low-resolution images

CUDA_VISIBLE_DEVICE=0 python predict.py --test_mode down --name down_x4 --scale 4 --resume_down ./experiment/ckpt/scale_4x/model_down_0110.pth --patch_size_down 512 --test_range 1-2000 --gpu cuda:6