bobpxx/lensless_imaging_transformer

Title: Image reconstruction with Transformer for mask-based lensless imaging

Authors: Xiuxi Pan, Xiao Chen, Saori Takeyama, Masahiro Yamaguchi

DOI: https://doi.org/10.1364/OL.455378

ol-47-7-1843.pdf is the paper.

News Reports related to this paper:

WIRED, Nikkei, Phys.org, EurekAlert!, Tokyo Tech News, et al.

Awards related to this paper:

2022年度（令和4年度）手島精一記念研究賞,

第38回 (2022年度) 電気通信普及財団賞

MESSAGE

Lensless Camera based on this paper has been productized and commercialized. Here are some potential applications:

1. Replace traditional camera in scenarios where space, weight or cost is extremely imposed, e.g.,
   • cost-sensitive IoT devices,
   • under-screen camera,
   • a space that is too limited for placing a traditional camera.
   • ...
2. Invisible spectrum (e.g, gama-ray, X-ray) imaging
   • Invisible spectrum imaging is too expensive or impossible for traditional lensed camera because of the usage of lens.
3. Optics-level privacy-preserving and cryptographic imaging/sensing
   • The captured encoded pattern is uninterpretable for human. We take this feature to develop privacy pretection and encryption.
   • My another project reconstruction-free lensless sensing verified that direct object recognition on the uninterpretable encoded pattern is possible.

You are warmly welcome to join me for further production development or extended research. You are also welcome for any question or discussion. Please contact me through My LinkedIn homepage or email.

DATASET

The datasets are available in Yamaguchi Lab OneDrive (It may be a temporary place, we are trying to seek a permanent place if many people are interested in it.)

There are three datasets:

1. mirflickr25k

• encoded pattern:
  • mirflickr25k_1600.zip in the OneDrive
• original images:
  • available in this link
• pattern-image matchup:
  • the corresponding encoded pattern and original image have the same name, only different in file extension. e.g., pattern "im1.npy"<-> image "im1.jpg".

2. dogs-vs-cats

• encoded pattern:
  • PetImages_1600.zip in the OneDrive
• original images:
  • available in this link, only 25k images in train folder are used.
• pattern-image matchup:
  • encoded patterns of dog & cat are separated to different folders. e.g., pattern "Cat/0.npy"<-> image "cat.0.jpg", pattern "Dog/1965.npy"<-> image "dog.1965.jpg"

3. fruits

• encoded pattern:
  • fruits_modified.zip in the OneDrive
• original images:
  • fruits_modifiedori.zip in the OneDrive
• pattern-image matchup:
  • same name, only different in file extension. e.g., pattern "n07739125_7447.npy"<-> image "n07739125_7447.JPEG".

The data collection method is written in page 3 of the original paper. The program to control the sensor for data collection is available in my another repository.

USAGE

Training

datasets/prepare_datasets.py prepares .npy files of dataset address;

configs.yaml defines training implementations;

train.py stars training.

An example of running train.py in linux:

CUDA_VISIBLE_DEVICES=0,1 nohup python -m torch.distributed.launch --nproc_per_node=1 --master_port 29501 train.py &

Prediction

predict.py starts prediction.

Checkpoint (checkpoints/best.pth) and input patterns (result/in-wild/pattern/ and result/on-screen/pattern/) can be used to verify our results.

Note

GrayPSF.npy is PSF of our lensless camera. It is not used in this reconstruction method, but a useful file to evaluate status of the optical system.

lensless_imaging_transformer's Issues

Can you provide the environmental setting?

Great job! Could you please share the environmental setting for this work? I am trying to implement this code on another dataset. Thanks!

License

Hi thanks for your open-sourcing your code! I was wondering if you could add a license to your code? So that it is clear how we can use, copy, distribute, or modify it. Thanks!

Hello, may I ask how to use this code. Looking forward to your reply！