How pksvision/deep-wavenet-underwater-image-restoration works
Main components
a. External Datasets & Resources
• EUVP test/train folders
• UIEB & Challenging‐60
• UFO‐120 (2X/3X/4X)
• Sample degraded videos (underwater_video.mp4)
• resources/existing_soas_list_with_codes.txt
b. Configuration & Installation
• installation/requirements.txt
• options.py files in each task module
c. Core Modules (per task)
• uie_euvp/
• uie_uieb/
• super-resolution/2X, 3X, 4X
• uw_video_processing/
d. Shared Utilities
• utils/gen_seg_maps.py
• matlab_iqms/ and python_iqms/ for SSIM/PSNR/UIQM
• loss‐plotting & other misc scripts
e. DeepWaveNet Model Definition
• models.py in each module (network architecture with WaveNet blocks)
• vgg.py (for perceptual loss)
f. Experiment Engines
• train.py (orchestrates data loading, model, loss, optimizer, checkpointing)
• test.py (loads checkpoint, runs inference, saves outputs)
g. Outputs & Evaluation
• Checkpoints in ckpts/
• Enhanced / super‐resolved images in output directories
• Metrics (SSIM, PSNR, UIQM) computed via utils
Relationships & data flows
• External Datasets → Dataset.py → DataLoader
• DataLoader → train.py / test.py → Model (models.py + vgg.py)
• Model → Restored Images → save to disk
• After inference → utils/python_iqms or utils/matlab_iqms → Metrics output
• train.py saves checkpoints → test.py loads checkpoints
• uw_video_processing/test.py uses the same model APIs for frame‐by‐frame processing
Architectural patterns & design principles
• Modular separation by task (UIE vs. SR vs. Video)
• Reusable Dataset / Model / Options components across tasks
• Configuration‐driven experiments (options.py)
• Checkpointing for reproducibility
• Shared utility library for metrics and post‐processing
Key technologies & frameworks
• Python 3.x
• PyTorch (deep learning)
• torchvision, numpy, scipy, opencv
• Jupyter (DeepWaveNet_demo.ipynb)
• Optional: MATLAB for IQM computations
• Deployment mention: Heroku web‐app (256×256 demo)