Burst Denoising with Kernel Prediction Networks
Ben Mildenhall1* Jon Barron2 Jiawen Chen2 Dillon Sharlet2 Ren Ng1 Robert Carroll2
1UC Berkeley      2Google
* Work done while interning at Google.

A qualitative evaluation of our model on real image bursts from a handheld camera in a low-light environment. The reference frame from the input burst (a) is sharp, but noisy. Noise can be reduced by simply averaging (b) a burst of similar images, but this can fail in the presence of motion. Our approach (f) learns to use the information present in the entire burst to denoise a single frame, producing lower noise and avoiding artifacts compared to baseline techniques (c - e).

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.

@inproceedings{mildenhall2018kpn,   author = {Mildenhall, Ben and Barron, Jonathan T and Chen, Jiawen and Sharlet, Dillon and Ng, Ren and Carroll, Robert},   title = {Burst Denoising with Kernel Prediction Networks},   booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},   year = {2018} }