Dual-domain image denoising

Abstract

Image denoising is a classic problem in digital image processing, which is half a century old. Despite its age, it remains an active subject. State-of-the-art image denoising methods produce objectively excellent results and approach theoretical limits. However, they still suffer from visually disturbing artifacts. While they produce acceptable results for natural images, human eyes are less forgiving when viewing synthetic images. Their success in denoising natural images is believed to be based on the use of patches, and the ability to learn statistics about them. Unfortunately, patch-based algorithms producing high-quality results are sophisticated and difficult to implement. We present a class of new image denoising algorithms that does not use patches and yet produces high-quality images. Our algorithms produce competitive results for grayscale images and surpass the state-of-the-art for color images. Unlike other state-of-the-art methods, the methods presented in this thesis produce nearly artifact free images, making them suitable for denoising synthetic images. Our methods are based on a new filter which operates in the spatial domain and in the frequency domain. This new filter is versatile and can also be used for removing denoising and compression artifacts. Our results demonstrate that patches are not essential to image denoising. Without the additional complexity of patches, our algorithms are surprisingly simple.

Data Adaptive Dual Domain Denoising: a Method to

“…The DDID algorithm [15] adapts the denoising patches (or blocks) to the shapes of the guide image and denoises them via Fourier shrinkage. The original algorithm iterated the process to obtain a guide image for the next step improving the results over consecutive iterations.…”

“…Several methods [14,16,27] use a single step of DDID with a guide image produced by a different algorithm. This yields much better results than the original DDID.…”

“…DDID [15] is an iterative algorithm that uses a guide image (from a previous iteration) to de-termine spatially uniform regions to which Fourier shrinkage could be applied without introducing ringing artifacts. Several methods [14,16,27] use a single step of DDID with a guide image produced by a different algorithm.…”

Optimizing the Data Adaptive Dual Domain Denoising Algorithm

“…This allows one to mix different denoising principles, thus yielding high quality results [26,2]. DDID [9] is an iterative algorithm that uses a guide image (from a previous iteration) to determine spatially uniform regions to which Fourier shrinkage could be applied without introducing ringing artifacts. Several methods [8,10,20] use a single step of DDID with a guide image produced by a different algorithm.…”

“…DDID [9] is an iterative algorithm that uses a guide image (from a previous iteration) to determine spatially uniform regions to which Fourier shrinkage could be applied without introducing ringing artifacts. Several methods [8,10,20] use a single step of DDID with a guide image produced by a different algorithm. This yields much better results than the original DDID.…”