Image Denoising Autoencoder

The need

Digital photos taken in low light carry noise from the camera sensor, and it depends on the sensor, its electronics and the conditions. Many denoising models train on artificial noise and struggle on real images. This project uses real noise, with clean reference shots of the same scene.

What I owned

The data protocol, model design, training and evaluation, in a team of two.

The solution

Convolutional autoencoder with encoder and decoder blocks — The convolutional autoencoder, with encoder and decoder blocks around a compressed code.

Data

RENOIR Canon T3iSplit by scene: 70/15/15

Preprocess

RGB in [0, 1]128×128 overlapping patches

Train

MSE lossReduceLROnPlateauPSNR and SSIM tracked

Evaluate

Overlap-tiled full imagesResidual histogramsOther sensor test

The data protocol stayed the same across experiments, so comparisons are fair.

How it was built

The split is made by scene and not by image, because several noisy shots share one reference and mixing them across sets would leak information. Evaluation reconstructs full images by tiling overlapping patches, which measures what a user would actually see.

Results

On the test images: a mean PSNR of 32.45 dB (±1.93) and a mean SSIM of 0.886.
On a different sensor (Mi3): 27.56 dB and 0.80 across 76 images. The drop is expected, and the metrics stay acceptable, which shows partial generalisation.

Dim scene with sensor noise — Noisy input.

The same scene after denoising — Denoised output.

Histogram showing denoised residuals more tightly centred on zero than noisy residuals — Residuals against the reference: denoised values cluster more tightly around zero.