DACAR-SACAR models

Artifacts reduction in JPEG-Compressed Images using CNNs.

DACAR Network/ SACAR Network

In this work, we consider both types of CNN architectures to reduce the artifacts (e.g.; blurring, blocking artifacts and ringing artifacts) in JPEG compressed images. We have proposed two CNNs which we refer to as Direct Architecture Compression Artifacts Removal (DA-CAR) network and Skip Architecture Compression Artifacts Removal (SA-CAR).

• For DA-CAR version, we experiment with 3, 4, and 5 layer architectures (DACAR3, DA-CAR4 and DA-CAR5, respectively).
• For the skip-based architecture we consider 6 layers (SACAR6), the third layer is concatenating activation (feature maps) between the first layer and the second layer (2+1).

Direct and Skip Architectures.

Network-Keras

These codes are to reduce the different artifacts from JPEG compressed images.

If these codes are helpful for you, please cite this paper: Artifacts reduction in JPEG-Compressed Images using CNNs, F. Albluwi, V. Krylov and R. Dahyot Irish Machine Vision and Image Processing conference (IMVIP 2018 https://www.ulster.ac.uk/conference/imvip-2018), 2018. Published in IMVIP e-book of proceedings with ISBN 978-0-9934207-3-3.

Dependencies

1. Python 3.6.5
2. TensorFlow 1.1.0.
3. Keras 2.2.2.
4. Matlab.
5. Matconvnet.

Generating data

1. Reduce the quality of images at different levels (JPEG_Quality = 10 or 20) by using 'Quality' function in Matlab.
2. The training set is 400 images from the BSDS500 (The Berkeley Segmentation Dataset).

Training

1. Generate training patches using Matlab: run generate_train.m and generate_test.m which in train folder, and then put this folder in the network folder you want to train (._train as SA-CAR6_train).
2. Use Keras with TensorFlow (tf) as a backend to train any model (DACAR3, DACAR4, DACAR5 or SA-CAR6); Adam is used to optimizing the network for fast convergence: run DACAR_train.py or SACAR_train.py to produce DACAR model / or SACAR model.
3. Convert Keras model to .Mat for testing using Matconvnet: run load_save.py first, then run save_model.m to produce Matconvnet model, and then put the .mat file in models folder which in test folder.
4. Run DACAR_SACAR_test.m in “test” folder to test the model; Live1 (which contains 29 images) and BSD100 (which contains 100 images) are used as testing data.

Some Qualitative Results

Qualitative evaluation of reconstruction quality using different networks for JPEG quality quality = 10 and quality =20.

The Quantitative Results