Simply Blur Detector

Original Dockerfile is tested on x86/64.

If your computer is using Apple M1 series chips, please use Dockerfile.M1 tested on my MBP.

Simply blur detector model is trained with 3 layers CNN by using Tensorflow 2.7.

It can be used to distinguish blur and non blur images.

Blur Generator can be found in https://github.com/NatLee/Blur-Generator

Usage

Run the demo, just use bash start.sh.

Or you can replace with your own non-blur images in folders ./data/test and ./data/train and run ./generate-blur-img.py.

If you want to fast run all steps, you can use the script ./fast-run.sh or ./fast-run.bat.

Training & Predicting

All materials are from the source https://bluearchive.wiki/wiki/Characters

• Training

  There is a function which can get image data from folder in ./src/blur_detection/utils/data_tools.py.

  I suggest that you should use Docker to train to model.

  from blur_detection import DetectionModel
  from blur_detection.utils.data_tools import get_data
  
  detector = DetectionModel() # without pre-trained model
  train_d, test_d = get_data('/data')
  detector.train(train_d, test_d)
  detector.save('/model/model.h5')

  The folder of data tree is shown as following.

  data
  |____train
  | |____blur
  | | |_____images
  | |____nonblur
  | | |_____images
  |____test
  | |____blur
  | | |_____images
  | |____nonblur
  | | |_____images

• Predicting

  If you want to predict an image and get a cool compare one like the image on the top of this document, just use the following script:

  from blur_detection import DetectionModel
  from blur_detection.utils import predict_blur_img
  
  detector = DetectionModel('<YOUR_MODEL_PATH>')
  predict_blur_img(
      detector,
      '<YOUR_IMG_PATH>',
      save_fig_path='<YOUR_FIG_SAVE_PATH>',
      blur_mode='<None|lens|gaussian|motion>' # 4 types, default is gaussian
  )

Author

Nat Lee