Burst Denoising with Kernel Prediction Networks

TL;DR

This paper introduces a neural network-based method for denoising bursts of images from handheld cameras, using spatially adaptive kernels, realistic synthetic data, and an annealed loss to improve performance across various noise levels.

Contribution

It proposes a novel convolutional neural network architecture for joint alignment and denoising of image bursts, along with a realistic data synthesis method and an optimization strategy to enhance denoising quality.

Findings

Outperforms state-of-the-art denoising methods on real and synthetic data.

Effectively handles a wide range of noise levels.

Achieves comparable or better results than existing techniques.

Abstract

We present a technique for jointly denoising bursts of images taken from a handheld camera. In particular, we propose a convolutional neural network architecture for predicting spatially varying kernels that can both align and denoise frames, a synthetic data generation approach based on a realistic noise formation model, and an optimization guided by an annealed loss function to avoid undesirable local minima. Our model matches or outperforms the state-of-the-art across a wide range of noise levels on both real and synthetic data.