Few methods for JPEG Image Compression Detection And Quality Factor Detection

Low Compression	Medium Compression	High Compression

Image taken from here

• Split the image into 8x8 blocks.
• Compute the DCT coefficients for each block and quantize the frequency components.
• Compute the normalized frequencies across all the blocks.
• Compute the variance of the high frequency blocks.
• If the variance is high, then unclamped high frequency components are present. If the variance is low, then jpeg quanitzed high frequency components are present.
• The above observation can be used to classify if a jpeg image is compressed.

Reconstructed Reference	Input	Residual

• Compute a reference denoised image using Real-ESRGAN.
• Compute the residual image, i.e., difference between reference denoised image and input image.
• The residual image contains high-frequency components, artifacts, etc.
• The residual image is passed to a resnet model to get classifcation and regression of jpeg-compression and quality factor respectively.
• The model is trained with augmenting single, doube jpeg compressions.
• Non-aligned double JPEG compression should be implicitly handled as the reference image should produce a constant noise pattern.

• We compute the JPEG Quality Factor using a pretrained regressor, Towards Flexible Blind JPEG Artifacts Removal (FBCNN, ICCV 2021) which supports single, doube jpeg compressions with Non-alignment of the quantized matrices in double jpeg compression.
• The FBCNN repository is used.

Run get_dataset.sh to download the datasets.

bash get_dataset.sh

Execute dct_validation.py to obtain the validation results stored as csv in outputs/dct.csv

python3 dct_validation.py

Execute dnn_validation.py to obtain the validation results stored as csv in outputs/dnn.csv

python3 dnn_validation.py

Execute fbcnn_validation.py to obtain the validation results stored as csv in outputs/fbcnn.csv

python3 fbcnn_validation.py

To train the DNN model, we run train.py

python3 train.py