Skip to content

google/attention-center

Attention center

This repository contains

  • a TensorFlow Lite model to that can be used to predict the attention center of an image, i.e. the area where the most salient parts of an image lie.
  • a python script that can be used to batch encode images using the attention centers. This can be used with the submodule libjxl in order to create JPEG XL images such that decoding the image will start from the attention center determined by the model.

Using Saliency in progressive JPEG XL images is a blog post from 2021 about progressive JPEG XL images.

Open sourcing the attention center model is a blog post from 2022 about this open source project.

For the training of the model center.tflite has been trained with images from the Common objects in context annotated with saliency from the salicon dataset.

How to use

Make sure you have python and tensorflow installed.

  1. Clone it from GitHub including the submodules
  2. Build libjxl, for example by following the instructions given in the the libjxl repo.
  3. run the encode_with_centers.py script.

Here ${INPUT_IMAGE_DIR} contains images that can be opened with the Python Imaging Library and ${OUTPUT_IMAGE_DIR} will be created if it does not exist and the encoded JPEG_XL images will be placed there.

git clone https://github.com/google/attention-center --recursive --shallow-submodules
cd attention-center/libjxl
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF ..
cmake --build . -- -j$(nproc)
cd ../../
python encode_with_centers.py --lite_model_file=./model/center.tflite \
  --image_dir="${INPUT_IMAGE_DIR}" --output_dir="${OUTPUT_IMAGE_DIR}"

There are the following flags:

  --[no]dry_run: If true, only do a dry run, do not write files.
    (default: 'false')
  --encoder: Location of th encoder binary.
  --image_dir: Name of the directory of input images.
    (default: './libjxl/build/tools/cjxl')
  --lite_model_file: Path to the corresponding TFLite model.
  --new_suffix: File extension of the compressed file.
    (default: 'jxl')
  --output_dir: Name of the directory of the output images.
  --[no]verbose: If true, prints info about the commands executed.
    (default: 'true')

An example for using -- would be

python encode_with_centers.py --lite_model_file=./model/center.tflite   --image_dir=/tmp/images --output_dir=/tmp/out/ -- -distance 1.1

Here we pass the flag --distance 1.1 to cjxl.

The flags and arguments for --center_x, --center_y and --group_order are automatically injected.

Example attention center calculations

Here we show an example image, where the calculated attention center of an image is computed. Running

python encode_with_centers.py --lite_model_file=./model/center.tflite --image_dir=./assets --dry_run

gives the following verbose output:

libjxl/build/tools/cjxl_ng -center_x 522 -center_y 1143 assets/white_eyes_bee.jpg white_eyes_bees.jpg

This tells us that the computed attention center is at pixel coordinates (522, 1143). We mark the attention center with a red dot and compare it with the original image.

original image image with attention center as red dot

Progressive JPEG XL demo in Chrome

Check out how some JPEG XL images encoded with the help of the attention center model look like at different stages when loaded with Chrome: google.github.io/attention-center


Authors: Moritz Firsching and Junfeng He

About

No description, website, or topics provided.

Resources

License

Code of conduct

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages