yuki-11 / csbsr

Official PyTorch implementation of the paper "Joint Learning of Blind Super-Resolution and Crack Segmentation for Realistic Degraded Images", IEEE Transactions on Instrumentation and Measurement (TIM) 2024. CSBSR is an advanced version of our previous work CSSR [MVA'21].

[IEEE TIM'24] CSBSR:
Joint Learning of Blind Super-Resolution and Crack Segmentation for Realistic Degraded Images

Yuki Kondo · Norimichi Ukita

Toyota Technological Institute
IEEE Transactions on Instrumentation and Measurement (TIM) 2024

Paper | Project Page

🚀 CSBSR [IEEE TIM'23] is the advanced version of CSSR [MVA'21]!

Crack segmentation challenges for synthetically-degraded images given by low resolution and anisotropic Gaussian blur. Our method (f) CSBSR succeeds in detecting cracks in the most detail compared to previous studies (d), (e). Furthermore, in several cases our method was able to detect cracks as successfully as when GT high-resolution images were used for segmentation (c), despite the fact that our method was inferring from degraded images.

---

Proposed joint learning network with blind SR and segmentation.

Table of Contents

• About the Project
• Usage
• License
• Citation
• Acknowledgement
• Contact

About the Project

This paper proposes crack segmentation augmented by super resolution (SR) with deep neural networks. In the proposed method, a SR network is jointly trained with a binary segmentation network in an end-to-end manner. This joint learning allows the SR network to be optimized for improving segmentation results. For realistic scenarios, the SR network is extended from non-blind to blind for processing a low-resolution image degraded by unknown blurs. The joint network is improved by our proposed two extra paths that further encourage the mutual optimization between SR and segmentation. Comparative experiments with State-of-the-Art segmentation methods demonstrate the superiority of our joint learning, and various ablation studies prove the effects of our contributions.

Usege

Setup Environment

git clone [email protected]:Yuki-11/CSBSR.git
cd CSBSR
virtualenv env
source env/bin/activate
pip install -r requirement.txt

Due to a temporary issue with wandb, specify the version of protobuf:

codepip install protobuf==3.20.1

Download Model Weights

Download the model weights from here and save them under weights/. The performance of the released models is as follows.

Model	IoUmax	AIU	HD95min	AHD95	PSNR	SSIM	Link
CSBSR w/ PSPNet (β=0.3)	0.573	0.552	20.92	22.52	28.75	0.703	here
CSBSR w/ HRNet+OCR (β=0.9)	0.553	0.534	17.54	20.29	27.66	0.668	here
CSBSR w/ CrackFormer (β=0.9)	0.469	0.443	39.37	56.59	25.93	0.571	here
CSBSR w/ U-Net (β=0.3)	0.530	0.506	26.33	27.24	28.68	0.702	here
CSSR w/ PSPNet (β=0.7)	0.557	0.539	21.20	24.74	28.35	0.656	here
CSBSR w/ PSPNet (β=0.3)+wF (mF=1)	0.573	0.551	18.73	21.7	28.73	0.702	here
CSBSR w/ PSPNet (β=0.3)+wF (mF=1)+BlurSkip	0.550	0.528	18.06	19.1	28.65	0.702	here

Prepare Dataset

Download the crack segmentation dataset by khanhha from here and extract it under datasets/. For more details, refer khanhha's repository.

Additionally, download the degraded original test set from here and extract it under datasets/crack_segmentation_dataset. Finally, the directory structure should look like this:

├── datasets
│   └── crack_segmentation_dataset
│       ├── images
│       ├── masks
│       ├── readme
│       ├── test
│       ├── test_blurred
│       └── train

Testing

Run the following command for testing:

python test.py weights/CSBSR_w_PSPNet_beta03 latest --sf_save_image

Training

Run the following command for training:

python train.py --config_file config/config_csbsr_pspnet.yaml

To resume training from a checkpoint, use the following command:

python train.py --config_file output/CSBSR/model_compe/CSBSR_w_PSPNet_beta03/config.yaml --resume_iter <resume iteration>

License

Distributed under the Apache-2.0 license License. Some of the code is based on the reference codes listed in the "Acknowledgements" section, and some of these reference codes have MIT license or Apache-2.0 license. Please, see LICENSE for more information.

Citation

If you find our models useful, please consider citing our paper!

@article{kondo2024csbsr,
  author={Kondo, Yuki and Ukita, Norimichi},
  journal={IEEE Transactions on Instrumentation and Measurement}, 
  title={Joint Learning of Blind Super-Resolution and Crack Segmentation for Realistic Degraded Images}, 
  year={2024},
  volume={73},
  number={},
  pages={1-16},
}

Acknowledgement

This implementation utilizes the following code. We deeply appreciate the authors for their open-source codes.

• Lextal/pspnet-pytorch
• Dootmaan/DSRL
• openseg-group/openseg.pytorch
• LouisNUST/CrackFormer-II
• lterzero/DBPN-Pytorch
• Yuki-11/KBPN
• vinceecws/SegNet_PyTorch
• khanhha/crack_segmentation
• vacancy/Synchronized-BatchNorm-PyTorch
• XPixelGroup/BasicSR
• google-deepmind/surface-distance
• hubutui/DiceLoss-PyTorch
• JunMa11/SegLossOdyssey
• NVIDIA/DeepLearningExamples
• xingyizhou/CenterNet
• YutaroOgawa/pytorch_advanced

Contact

If you have any questions, feedback, or suggestions regarding this project, feel free to reach out:

• Email: [email protected]

csbsr's Issues

how to build the test file?

The test file should consider a config.yaml, images and a model files which save the weights? in the Usege, why does it create a weights folder? python test.py weights/CSBSR_w_PSPNet_beta03/latest --sf_save_image

Thank you