JPEG Compression Detection and Quality Estimation

A simple demonstration of JPEG compression detection and quality estimation using machine learning.

Research Process

Dataset Creation

Feature Engineering

The aim is to produce a tensor of fixed-length representing an image of arbitrary size. The custom feature extraction process is the following.

Let $W$ and $H$ be the image width and height, respectively. Let $B = 8$ represent a block size. The process below is described for a single channel. In the case of multiple channels, e.g., for YCbCr color mode, the very same algorithm would be adopted separately.

1. Image padding. Assure that both image dimensions $W$ and $H$ are divisible by $B$, thus $\tilde{W} = W + \Delta w$ such that $\exists k \in \mathbb{N}$, such that $\tilde{W} = Bk$. The same applied for $\tilde{H}$. If necessary, expand the image by copying the edge values.
2. Block splitting. Split the image of size $\tilde{W} \times \tilde{H}$ into equal $B \times B$ blocks. Let $N$ denote the number of produced blocks.
3. Block reshaping. Merge all the $N$ extracted blocks into a single tensor of shape $N \times B \times B$.
4. Reduction. Apply $R$ different statistical reductions, each time using a different function, e.g., min, max, mean, standard deviation, or median. Each reduction will produce $R$ distinct $B \times B$ matrices.
5. Zig-zag selection and Concatenation. For subsequent visualization sake, a zig-zag selection is utilized as a substitute for a flatten operation (the order of indices does not affect ML algorithms). Thus, each of the $R$ matrices with shape $B \times B$ is converted into a single-dimensional vector of length $B^2$. The resulting vectors are concatenated to form the final feature vector of length $R \cdot B^2$.

There are several notable observations. Given the feature extraction strategy above, the contribution of the minimum and maximum statistics is the least significant. It is completely negligible. As a result, the model trained using just mean, standard deviation, and median performs just as well with considerably fewer parameters.

References

Datasets

• TIFF files dataset.

Relevant Research Papers

• Robinson, Jonathan, and Vojislav Kecman. "Combining support vector machine learning with the discrete cosine transform in image compression." IEEE Transactions on Neural Networks 14.4 (2003): 950-958.
• Retraint, Florent, and Cathel Zitzmann. "Quality factor estimation of jpeg images using a statistical model." Digital Signal Processing 103 (2020): 102759.