Skip to content

Latest commit

 

History

History
148 lines (115 loc) · 6.5 KB

README.md

File metadata and controls

148 lines (115 loc) · 6.5 KB
Raw

CLIQA

cliqa provides low-vision image quality scores that are more consistent across different images.

It is useful for filtering low-quality images with a threshold value when creating image datasets.

JPEGQuality

Content-based JPEG Quality predictor.

This can detect low-quality JPEG images even in the following cases.

  • (past) JPEG encoded image in PNG format
  • Re-encoded JPEG image with the higher quality value than the original quality value

The current pre-trained model can predict JPEG quality value around an average error of 3/100 on clean validation data. In real world data it should be much worse. However, I feel that it has achieved a practical level.

cliqa.filter_low_quality_jpeg

This tool copies only high quality images in the directory specified with -i to the directory specified with -o.

pytyon -m cliqa.filter_low_quality_jpeg -i /path_to_input_dir -o /path_to_output_dir

options.

options:
  -h, --help            show this help message and exit
  --input INPUT, -i INPUT
                        input image directory (default: None)
  --output OUTPUT, -o OUTPUT
                        output image directory (default: None)
  --checkpoint CHECKPOINT
                        model parameter file (default: ./cliqa/pretrained_models/jpeg_quality.pth)
  --gpu GPU [GPU ...], -g GPU [GPU ...]
                        GPU device ids. -1 for CPU (default: [0])
  --num-patches NUM_PATCHES
                        number of 128x128 patches used per image (default: 8)
  --quality QUALITY     quality threshold (default: 90)
  --filter-420          drop all 4:2:0(chroma-subsampled) JPEG images (default: False)
  --symlink             create symbolic links, instead of copying the real files (recommended on linux) (default: False)
  --score-prefix        add score prefix to the output filename (default: False)

The threshold value is specified with --quality option (default: 90, it is a predicted JPEG quality).

On Linux, it is recommended to always use --symlink option.

The predicted quality of each image can be added to the filename with --score-prefix option. Then, sorting by filename is in order of quality. --quality 0 --score-prefix allows the score prefix to be added to all files without filtering out low quality images.

GrainNoiseLevel

Content-based Photo Noise Level predictor. This can detect noisy photos.

The predicted value is noise_level = 50 - min(PSNR(original_image, noise_added_image), 50), so predicted_psnr = 50 - predicted_noise_level. predicted_psnr >= 40 is predicted to be a less noisy photo.

Note that this model is trained with not very perfected synthetic noise. May not work well in some real world noise cases.

cliqa.filter_noisy_photo

This tool copies only non-noisy images in the directory specified with -i to the directory specified with -o.

pytyon -m cliqa.filter_noisy_photo -i /path_to_input_dir -o /path_to_output_dir

options.

options:
  -h, --help            show this help message and exit
  --input INPUT, -i INPUT
                        input image directory (default: None)
  --output OUTPUT, -o OUTPUT
                        output image directory (default: None)
  --checkpoint CHECKPOINT
                        model parameter file (default: ./cliqa/pretrained_models/grain_noise_level.pth)
  --gpu GPU [GPU ...], -g GPU [GPU ...]
                        GPU device ids. -1 for CPU (default: [0])
  --num-patches NUM_PATCHES
                        number of 128x128 patches used per image (default: 8)
  --psnr PSNR           quality threshold (default: 40)
  --symlink             create symbolic links, instead of copying the real files (recommended on linux) (default: False)
  --score-prefix        add score prefix to the output filename (default: False)

The threshold value is specified with --psnr option (default: 40, it is a predicted PSNR).

Other options are the same as filter_low_quality_jpeg.

Note

If you want to use both filter_noisy_photo and filter_low_quality_jpeg, run filter_low_quality_jpeg first, then filter_noisy_photo. The reason for this is that filter_noisy_photo does not use JPEG quality less than 80 for training.

Example,

pytyon -m cliqa.filter_low_quality_jpeg -i ./data/images -o ./data/hq1 --symlink
pytyon -m cliqa.filter_noisy_photo -i ./data/hq1 -o ./data/hq2 --symlink

More practical example,

# rename original `images` directory to `images_original` temporary.
mv ./data/images ./data/images_original
# filtering low quality jpegs
pytyon -m cliqa.filter_low_quality_jpeg -i ./data/images_original -o ./data/images_filter_jpeg --symlink
# filtering noisy photos
pytyon -m cliqa.filter_noisy_photo -i ./data/images_filter_jpeg -o ./data/images_filter_noisy --symlink
# Make the original `images` directory link to the cleaned directory. (replace)
ln -s `realpath ./data/images_filter_noisy` ./data/images

ScaleFactor

Content-based upscaling factor predictor. This can detect upscaled images with interpolation filter.

The predicted value is quality = 100 - (predicted_scaler_factor - 1) * 100. predicted_scale_factor ranges from 1.0 to 2.0. 100 means the image is original scale or downscaled. 80 means the image is upscaled by 1.2x.

Note that this model is trained with very simple synthetic data. This can false positive detect blurs other than upscaling, such as depth of field effect or interpolation of rotation.

cliqa.filter_low_quality_resize

This tool copies only non-upscaled images in the directory specified with -i to the directory specified with -o.

pytyon -m cliqa.filter_low_quality_resize -i /path_to_input_dir -o /path_to_output_dir

options.

  -h, --help            show this help message and exit
  --input INPUT, -i INPUT
                        input image directory (default: None)
  --output OUTPUT, -o OUTPUT
                        output image directory (default: None)
  --checkpoint CHECKPOINT
                        model parameter file (default: ./cliqa/pretrained_models/scale_factor.pth)
  --gpu GPU [GPU ...], -g GPU [GPU ...]
                        GPU device ids. -1 for CPU (default: [0])
  --num-patches NUM_PATCHES
                        number of 128x128 patches used per image (default: 8)
  --quality QUALITY     quality threshold (default: 95)
  --invert              extract low quality resized images (default: False)
  --symlink             create symbolic links, instead of copying the real files (recommended on linux) (default: False)
  --score-prefix        add score prefix to the output filename (default: False)