Hardware Implementation of Low-latency Image Denoising with Wavelet Shrinkage and Noise2Noise-based Learning for Ultra-high Definition Cameras

There is a growing demand for real-time image denoising in low-light shooting with ultra-high definition cameras. This paper presents a denoising method that incorporates Haar-wavelet shrinkage denoising and a minimum mean square error estimation based on an extension of the Noise2Noise learning technique, which can perform low-latency processing and can be adapted to non-linear processed signals without prior knowledge of noise and signal distributions. By introducing a two-dimensional kernel to reduce memory usage in the coring function estimator, we demonstrate real-time hardware denoising for luminance signals of 4K 60 fps video with a delay of less than one millisecond.