Digital Image Processing

Electronic version: Digital Image Processing (Chinese) Third Edition (1).pdf

Study chapters 1, 2, 4, 6.1–6.2, and 8. Chapter 8 can be studied together with the JPEG Encoding Details Introduction, to understand the general encoding workflow.

Deep Learning

First, ensure sufficient proficiency in Python. Optional e-book: Python Programming: From Beginner to Practice.pdf

Learn PyTorch. Bilibili course: Li Mu's AI Learning Space - Hands-on Deep Learning (PyTorch Version) - Bilibili Video (bilibili.com), focus on 00 to 29.2, 31, 33–37, 47, 47.2

Foundational Papers for Beginners

Combine reading papers with code (CompressAI) to train your own model and plot RD curves.

Common training and validation datasets: ImageNet/COCO

Common test datasets: 24 Kodak images. Due to abnormal edges in original images, sometimes use cropped square versions.

1. Ballé, J., et al. (2015). "Density modeling of images using a generalized normalization transformation." arXiv preprint arXiv:1511.06281.

   The GDN activation layer commonly used in AI Codec. Related code: CompressAI/compressai/layers/gdn.py at master · InterDigitalInc/CompressAI (github.com)

2. Ballé, J., et al. (2016). "End-to-end optimized image compression." arXiv preprint arXiv:1611.01704.

   Introduces the foundational architecture of AI Codec. Can be studied alongside JPEG encoding, comparing shared steps such as transformation, quantization, entropy coding, and understanding the RD loss function. Related code: CompressAI/compressai/models/google.py at a4ae2eeef7bdb1b84ba076ac0d650b523f3fa882 · InterDigitalInc/CompressAI · GitHub

3. Ballé, J., et al. (2018). "Variational image compression with a scale hyperprior." arXiv preprint arXiv:1802.01436.

   Adds a hyperprior (hyper) to the base architecture. Related code: CompressAI/compressai/models/google.py at a4ae2eeef7bdb1b84ba076ac0d650b523f3fa882 · InterDigitalInc/CompressAI · GitHub

4. Minnen, D., et al. (2018). "Joint autoregressive and hierarchical priors for learned image compression." Advances in neural information processing systems.

   Autoregressive and hyperprior, related code: CompressAI/compressai/models/google.py at a4ae2eeef7bdb1b84ba076ac0d650b523f3fa882 · InterDigitalInc/CompressAI · GitHub

Note: CompressAI can be directly installed via pip on Linux, but no Windows package is provided. Follow the steps below for installation:

1. Install C++ packages in Visual Studio and restart.
2. Use conda to create a Python environment, then install pytorch, scipy, matplotlib, and pytorch-msssim.
3. Activate the environment, clone CompressAI locally, navigate to the directory in terminal, and run pip install .