A noise-adaptive CNN framework that classifies image noise type and routes inputs through specialized denoising branches, evaluated against clean ground-truth images across multiple benchmark datasets.
| Field | Details |
|---|---|
| Course Title | Current Topics in Machine Learning |
| Course Number | CAP 6614 |
| Semester | Fall 2025 |
| Assignment Title | Final Group Project |
| Assignment Description | Design and implement a novel approach or reproduce results from a current machine learning research paper. This project expands on adaptive denoising research by developing a noise-type-aware CNN that classifies the noise present in an input image and routes it through a specialized denoising branch, then compares this routing approach against a single comprehensive denoising model evaluated on clean ground-truth images. |
This project implements and compares two CNN-based image denoising strategies: a single comprehensive model trained across all noise types, and a routing model that first classifies the noise type then applies a noise-specific denoiser. Six synthetic noise types are supported (Gaussian, salt-and-pepper, uniform, Poisson, JPEG compression, and impulse), and each routing branch is architecturally tailored to its noise class. Models are trained and evaluated on MNIST, CIFAR-10, CIFAR-100, and STL-10 using a composite loss function combining MSE, MAE, SSIM, and gradient loss. The core research question is whether adaptive routing yields measurably better reconstruction quality than a single unified model when both are evaluated directly against clean ground-truth images.
No screenshot available. Add one with:
After training, each model is evaluated on a held-out test set and results are written to ./Results/. The routing system additionally produces a confusion matrix for the noise classifier.
Sample terminal output (comprehensive model, CIFAR-10):
============================================================
FINAL TEST RESULTS
============================================================
Test Loss: 0.0321
Test PSNR: 31.05 dB
Test SSIM: 0.9858
Total Computation Time: 00:43:12
============================================================
Sample terminal output (routing model, STL-10, per noise type):
Noise Type Loss PSNR (dB) SSIM
Gaussian 0.0400 29.54 0.9859
Salt & Pepper 0.0091 35.36 0.9974
Uniform 0.0240 32.84 0.9937
Poisson 0.0644 24.63 0.9599
JPEG 0.0448 27.95 0.9806
Impulse 0.0100 35.99 0.9971
Overall 0.0321 31.05 0.9858
Routing Accuracy: 99.66%
How to interpret the output:
- Loss is the composite loss (MSE 35% + MAE 35% + SSIM 20% + Gradient 10%). The project target is loss < 0.03 per noise type.
- PSNR (Peak Signal-to-Noise Ratio) is reported in dB; higher is better. Values above ~30 dB indicate good reconstruction quality.
- SSIM ranges from 0 to 1; values above 0.98 indicate near-perfect structural similarity to the clean image.
- Routing Accuracy reflects how often the classifier correctly identifies noise type before routing.
Key files produced per run:
./Results/
Comprehensive/
results_summary.csv # Aggregated metrics across all datasets
<dataset>/
training_log.txt # Epoch-by-epoch loss, PSNR, SSIM
results.json # Full results dictionary
epoch_data.csv # Per-epoch training/validation metrics
visual_comparisons/ # Side-by-side clean / noisy / denoised PNGs
images/ # Individual clean and denoised PNGs
Routing/
results_summary.csv
<dataset>/
confusion_matrix.png # Noise classifier confusion matrix
training_log.txt
results.json
epoch_data.csv
./Models/Saved/
comp_<dataset>.pth # Saved comprehensive model weights
classifier_<dataset>.pth # Saved noise classifier weights
<noisetype>_<dataset>.pth # Saved routing branch weights
If loss remains above 0.03: check the training log for early stopping triggers and consider increasing --epochs or adjusting --learning_rate. The Gaussian, Poisson, and JPEG branches are the most likely to need additional tuning.
image-denoising cnn adaptive-routing noise-classification composite-loss psnr ssim early-stopping multi-branch-architecture sobel-gradient-loss
- Language: Python 3.x
- Framework/SDK: PyTorch, torchvision
- Build System: pip / requirements.txt
project-root/
├── Main.py # Entry point; argument parsing, dataset prep, training orchestration
├── requirements.txt # Python dependencies
│
├── Models/
│ ├── CompModel.py # Comprehensive single-model denoiser architecture
│ ├── NoiseClassifier.py # Noise type classifier (6-class CNN)
│ ├── Saved/ # Saved .pth model weights (generated at runtime)
│ └── Routing_Models/
│ ├── GaussianModel.py # Dilated convolution denoiser for Gaussian noise
│ ├── SaltPepperModel.py # Spatial attention denoiser for salt-and-pepper noise
│ ├── UniformModel.py # Wide-channel dropout denoiser for uniform noise
│ ├── PoissonModel.py # Instance norm dual-branch denoiser for Poisson noise
│ ├── JPEGModel.py # Deblocking denoiser for JPEG compression artifacts
│ └── ImpulseModel.py # Corruption-mask denoiser for impulse noise
│
├── Utilities/
│ ├── Data_Loader.py # AdaptiveDataset wrapper for MNIST/CIFAR-10/CIFAR-100/STL-10
│ ├── Noise_Generator.py # On-the-fly noise synthesis for all 6 noise types
│ ├── Composite_Loss.py # MSE + MAE + SSIM + Gradient composite loss function
│ ├── Comp_Train.py # Training loop for the comprehensive model (PSNR/SSIM tracking)
│ └── Class_Train.py # Training loop for the noise classifier (accuracy + confusion matrix)
│
├── Data/ # Downloaded datasets (created at runtime)
├── Results/ # Training logs, CSVs, PNGs (created at runtime)
│ ├── Comprehensive/
│ └── Routing/
└── Noise_Examples/ # Example noise visualizations (generated optionally)
- Python 3.8+
- CUDA-capable GPU (recommended; CPU fallback is supported)
# 1. Clone the repository
git clone https://github.com/alexneilgreen/UCF-CurrentTopicsInML-AdaptiveNoiseTypeRoutingCNN.git
cd UCF-CurrentTopicsInML-AdaptiveNoiseTypeRoutingCNN
# 2. Install dependencies
pip install -r requirements.txt
# 3a. Train the comprehensive model on CIFAR-10 (default)
python Main.py --model comp --dataset cifar10
# 3b. Train the routing model (classifier + all noise-specific branches) on CIFAR-10
python Main.py --model routing --dataset cifar10
# 3c. Train on all datasets
python Main.py --model comp --dataset all
python Main.py --model routing --dataset all| Argument | Default | Description |
|---|---|---|
--model |
comp |
Model type: comp (comprehensive) or routing (adaptive routing) |
--dataset |
cifar10 |
Dataset: mnist, cifar10, cifar100, stl10, or all |
--epochs |
50 |
Training epochs for denoiser models |
--classifier_epochs |
30 |
Training epochs for the noise classifier |
--learning_rate |
0.001 |
Adam optimizer learning rate |
--batch_size |
0 |
Batch size (0 = auto-adjust per dataset) |
--ESP |
10 |
Early stopping patience in epochs |
--output_dir |
./Results |
Root directory for all saved outputs |
--data_root |
./Data |
Root directory for dataset downloads |
--device |
auto | cuda or cpu |
--seed |
42 |
Random seed for reproducibility |
| Name | Role | GitHub |
|---|---|---|
| Alexander Green | Organization, Development, Testing and Presentation | @alexneilgreen |
This repository is publicly available for portfolio and reference purposes only. Please do not submit any part of this work as your own for academic coursework.