Low-Light Image Denoising using DnCNN

Every smartphone owner knows the problem: a photo taken in dim light comes out grainy, and the phone's built-in cleanup often smears fine detail away along with the noise. Classical filters like median or bilateral use fixed local rules — they cannot tell texture from grain, so they blur edges while removing noise. In 2017, Zhang et al. showed a cleaner way with DnCNN: instead of predicting the clean image directly, the network learns to predict the noise, which is then subtracted from the input. Residual learning is an easier problem — noise has simpler structure than natural images — and batch normalization plus a deep stack of small convolutions keeps training stable. One conditioned model covers a range of noise levels, exactly what real low-light photos need. This build implements DnCNN in PyTorch with a Flask demo: upload a noisy photo, set the noise level with a slider, and compare noisy and cleaned versions side by side. An evaluation notebook computes PSNR and SSIM on BSD68 during the build — a genuine, auditable experiment for the report.

Dataset & model:
Dataset names: BSD68 (68 natural grayscale images) and Set12 (12 images) — the standard denoising benchmarks used in the original paper. Source: the public Berkeley segmentation dataset family releases. Training uses clean natural images corrupted with synthetic Gaussian noise at several noise levels to form noisy/clean pairs. Task: image denoising (image-to-image regression). Classes: none — not a classification task.
Model: DnCNN — "Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising" (Zhang et al., arXiv:1608.03981, TIP 2017). Deep stack of 3x3 convolutions with batch normalization and ReLU, conditioned on noise level; the network predicts the noise residual, which is subtracted from the input.
Input: a noisy low-light photo (grayscale or color, JPEG/PNG, any size via patch stitching). Prediction: the noise residual per patch. Output: the cleaned photo, shown side by side with the original and downloadable.
Evaluation metrics: PSNR and SSIM on BSD68, computed by the evaluation notebook during the build; no scores claimed in advance. The report discusses the gap between synthetic training noise and real smartphone noise explicitly.

Working:

1. Dataset: clean natural images are corrupted with synthetic Gaussian noise at several noise levels to create noisy/clean training pairs; BSD68 and Set12 are reserved as evaluation benchmarks.
2. Preprocessing: images are split into overlapping patches so arbitrary image sizes can be processed; patches are normalized before being fed to the network.
3. Training phase: the DnCNN — a deep stack of 3x3 convolutions with batch normalization and ReLU — is trained to predict the noise residual for each patch, conditioned on the noise level; checkpoints are validated with PSNR on held-out data.
4. Evaluation phase (offline): the evaluation notebook runs the build's own weights over BSD68 and logs PSNR and SSIM across noise levels — no scores claimed in advance.
5. Inference phase (demo): an uploaded noisy photo is split into overlapping patches; each patch passes through the DnCNN, which outputs a predicted noise residual; the residual is subtracted from the input patch to give the clean estimate; patches are stitched back with overlap blending to avoid seams.
6. Output: the noise-level slider selects the model conditioning matching the photo's grain (light, medium, heavy presets), and the cleaned result is shown side by side with the original and is downloadable.

Specifications:
Model | DnCNN residual CNN (Zhang et al., arXiv:1608.03981, TIP 2017)
Core idea | Residual learning: the network predicts the noise, which is subtracted from the input
Architecture | Deep stack of 3x3 convolutions with batch normalization and ReLU
Datasets | BSD68 (68 images) and Set12 (12 images) for evaluation
Input | Noisy low-light grayscale or color photos, any size via patch stitching
Control | Noise-level slider (light/medium/heavy presets) selecting model conditioning
Output | Cleaned photo shown side by side with the original, downloadable
Demo app | Flask with upload, noise-level slider, side-by-side view and download
Evaluation | PSNR/SSIM on BSD68 computed by the notebook during the build; no scores claimed in advance
Weights | Trained DnCNN weights shipped with the build

DnCNN Residual-Learning Denoiser [implemented] — Predicts the noise residual and subtracts it, rather than predicting the clean image directly — the architecture from Zhang et al.
Noise-Level Slider [implemented] — Match the denoiser to the photo's grain, with light, medium and heavy presets plus fine control.
Grayscale and Color Support [implemented] — Both are handled in the same demo, with color channels processed jointly to avoid hue shifts.
Side-by-Side Comparison View [implemented] — Noisy and denoised versions are shown together so the improvement is directly visible.
BSD68 Batch Test Mode [implemented] — Run the whole BSD68 set through the pipeline via the notebook for benchmark-style evaluation.
PSNR/SSIM Evaluation Notebook [implemented] — Computes both metrics across noise levels during the build.
Patch-Based Inference [implemented] — Overlapping patches with blend stitching handle arbitrary image sizes without seams.
Blind (Unknown-Noise) Denoising [future-scope] — Automatic noise-level estimation without the slider is documented as a possible extension, not included.

Complete source code: DnCNN model, training loop, patch inference, Flask demo app
Trained DnCNN weights
Training notebook and evaluation notebook (PSNR/SSIM on BSD68)
Project report PDF: denoising background, residual-learning derivation, evaluation procedure, error analysis (synthetic-vs-real noise gap)
PPT presentation for final review
Viva Q&A preparation document: residual learning, batch normalization, noise models, PSNR/SSIM
Setup guide: environment, dependencies, GPU/CPU guidance

The model assumes roughly Gaussian noise; heavy JPEG compression artifacts or motion blur are different problems needing different methods.
An almost-black image has no signal to recover — denoising removes noise, it does not invent detail that was never captured.
High noise-level settings can over-smooth fine texture, trading grain for softness.
Real smartphone noise differs from the synthetic training noise, so real captures are the tougher test; the report discusses this gap explicitly.
This is an educational prototype, not a replacement for a camera's image pipeline.