How jzyustc/mbrs works
Project Type and Purpose
• This is a research‐oriented deep learning project (MBRS) focused on robust watermark embedding and extraction in images using PyTorch.
• It implements a mini‐batch strategy combining real and simulated JPEG compression during training to improve watermark robustness.
High‐Level Architecture Overview
The system can be broken down into three major layers:
A. Data & Configuration Layer
B. Core ML Pipeline (Network + Noise Simulation + Training/Test Loops)
C. Persistence & Results Layer
Main Components
1. Data & Configuration
• Datasets directory (ImageNet/COCO folders under datasets/)
• utils/ modules:
• Dataloader.py – wraps PyTorch DataLoader for training/test sets
• load_train_setting.py / load_test_setting.py – parse JSON hyperparameters
• settings.py – default constants
• Configuration files:
• train_settings.json
• test_settings.json
2. Core ML Pipeline
a. Network Module (network/)
• Encoder classes (Encoder_MP.py, Encoder_MP_Decoder.py)
• Decoder.py – extracts watermark bits
• Discriminator.py – adversarial discriminator for image realism
• Network.py – orchestrates the end‐to‐end forward pass
• Blocks submodule (network/blocks/): reusable layers (ConvNet, ExpandNet, SENet)
b. Noise Simulation (network/noise_layers/)
• Combined, crop, gaussian_filter, jpeg, salt_pepper, etc.
• Injects distortion between encoder and decoder during training
c. Training & Testing Scripts
• train.py: loads settings, instantiates DataLoader, network, noise layers, optimizer, loss functions, and executes the training loop with early stopping, checkpoint saving.
• test.py: loads a saved model checkpoint, runs inference over test images, records BER & PSNR metrics.
3. Persistence & Results
• results/: hierarchical storage per experiment (e.g., MBRS_256_m256/), containing:
• models/ – saved checkpoints
• images/ – sample reconstructions
• train_log.txt, val_log.txt, test_*.log, *.json – performance logs and parameters
Relationships & Data Flow
1. Configuration files → JSON parser (utils/load_*_setting.py) → training/test parameters
2. Dataset directory → utils/Dataloader.py → minibatches of images & watermark bits
3. minibatch → Encoder → watermarked images
4. watermarked images → “real + simulated JPEG” noise layers → corrupted images
5. corrupted images → Decoder → recovered watermark bits
6. (During training) Encoder+Decoder outputs → loss functions (+ Discriminator) → backprop through Network
7. Every N epochs → checkpoint saved into results/.../models/ → performance logged
Architectural Patterns & Design Principles
• Modular design:
• Separation of concerns between network layers (network/blocks), noise simulation (network/noise_layers), and orchestration (network/Network.py).
• Configurable pipeline: JSON‐driven hyperparameters allow reproducibility and multiple experiment configurations.
• Pluggable noise modules: easy to add new attacks by dropping new Python modules in noise_layers/.
• Clear logging and checkpointing for rapid iteration.
Technologies & Frameworks to Highlight
• PyTorch for model definition, training loops, autograd
• Kornia for differentiable image processing (when applicable)
• NumPy, Pillow, SciPy for data manipulation
• Python standard libs for JSON, logging, file I/O