Skoltech Professor Stamatios Lefkimmiatis has invented a new image denoising technology.
Image denoising plays a key role both in image processing and computer vision. The goal in image denoising is to remove “noise,” or visual distortions, from a captured image while preserving the original image’s important low-level features, such as edges or/and texture.
There are many different approaches one can take to denoise images. Among the most popular ones is to recover the original image as the solution of a large-scale optimization problem involving millions of unknown parameters. While this strategy has proven to be quite effective, it has two main drawbacks that hinder its wider utility for commercial applications.
First, this approach requires users to explicitly design a proper objective function, the minimization of which leads to the desired outcome. Second, users must numerically compute the minimizer of this objective. Designing the proper cost function so as to get a meaningful (in a physical or statistical sense) denoising result is by no means a trivial task, since different cost functions lead to recovered images of different image qualities. Moreover finding a numerical solution of the minimization problem is computationally very expensive.
Professor Lefkimmiatis’ new approach is based on neural network deep learning algorithms. He and his team designed a deep network that has proven highly efficient for image denoising of both gray-scale and color images.
Specifically, Professor Lefkimmiatis and his team proposed training a network to automatically learn the objective function in a supervised manner, in turn leading to optimal denoising results based on an image quality criterion of the user’s choice.
Not only does the new network learn the objective function; it also internally imitates the optimization steps necessary to obtain the desired solution. Network training can be a time-consuming process; the advantage of this approach however is that its execution after this phase is comparatively fast and produces excellent results.
Moreover, this method helps users avoid the necessity of making specific decisions about the form of the objective function to be minimized by enabling the network to learn from training examples.
“In our work we relied on recent advances in deep learning and we designed a novel network architecture that addresses very efficiently the drawbacks of the previous image denoising techniques and leads to state-of-the-art results in a relatively short amount of time,” said Professor Lefkimmiatis.
“The core component in the design of our network that makes it ideal for image denoising applications – and possibly for other related problems like image deblurring, super-resolution, inpainting, etc. – is the explicit modeling of the non-local self-similarity property of natural images. This property implies that images typically consist of patterns that are repeated in different and possibly distant image locations. This property highlights a redundancy of information that has been used in the past by standard non-machine-learning techniques, and has led to significant improvements in image restoration quality,” explained Professor Lefkimmiatis.
The novelty in this work is that scientists designed the first deep-network that has the ability to exploit the self-similarity property so as to distinguish between noise and low-level image structures and separate them in the most efficient way possible.
The results of the study will be presented at IEEE’s Conference on Computer Vision and Pattern Recognition, a highly competitive and influential venue that attracts the world’s leading computer vision researchers.
+7 (495) 280 14 81