How akshat8102005/denoising-images works
This repository implements an image-denoising workflow: training loads paired noisy/clean NumPy images, fits a residual CNN, and saves its best checkpoint; evaluation loads a checkpoint, denoises test images, and reports PSNR/SSIM while saving predictions and metrics. A separate baseline CNN is present, while README-listed future architectures and LPIPS comparison are roadmap items, not evidenced runtime components. Edges are limited to relationships shown in the sampled source or README; external dataset locations are represented as actors.
Main entry points: evaluate.py Inference and evaluation · train.py Model training · dataset.py Paired dataset loader · checkpoints Model checkpoint
Components 12
Data and training
- Paired dataset loader
datasets/dataset.py - Model training
train.py - Model checkpoint
results/checkpoints
Denoising models
- Residual CNN
models/residual_cnn.py - Baseline CNN
models/baseline_cnn.py
Evaluation
- Inference and evaluation
evaluate.py - PSNR metric
metrics/psnr.py - SSIM metric
metrics/ssim.py - Saved predictions
- Evaluation report
results/residual/metrics.txt
Other components
- Paired image files
- Test image files
Connections 11
- Paired image files → Paired dataset loader: provides pairs
datasets/dataset.py - Model training → Paired dataset loader: loads batches
train.py - Model training → Residual CNN: trains
train.py - Model training → Model checkpoint: saves best weights
train.py - Test image files → Inference and evaluation: provides images
evaluate.py - Inference and evaluation → Residual CNN: runs inference
evaluate.py - Inference and evaluation → Model checkpoint: loads weights
evaluate.py - Inference and evaluation → PSNR metric: calculates
evaluate.py - Inference and evaluation → SSIM metric: calculates
evaluate.py - Inference and evaluation → Saved predictions: writes predictions
evaluate.py - Inference and evaluation → Evaluation report: writes averages
evaluate.py