With a simple gradient interruption operation, the denoising network eventually converges to the “noise-to-clean” objective, while the noise extraction network learns to finely extract the noise. The denoising network is then trained using a new noisy/clean image pair obtained by adding the extracted noise to a random clean image (“noise-to-clean”). We use the noisy/blurred image pair to guide the noise extraction network to roughly extract noise from its input (“noise-to-blur”). The noisy inputs are first transformed into noise-free but blurred images by general filtering techniques ( e.g. Our N2B model consists of two subnetworks, i.e. Although the images we have are unpaired, we can extract information from them to generate supervision for the denoising process. It only requires some unpaired noisy images and clean images, which is easy to implement in most practical applications. The training of N2B model does not need access to estimation of noise and pre-collected paired data. In this paper, we propose Noise2Blur (N2B), a training scheme that overcomes the above problems. Gaussian noise, Poisson noise) and generalize poorly to the real-world noisy images with more complicated noise. As a result, CNN denoisers are easily over-fitted to simple synthetic noise ( e.g. In addition, the real noise degradation process is usually complex or unknown, so that the synthesized noise distribution can deviate severely from the real noise distribution. Since pairs of real noisy/clean images are often impossible to obtain, CNN denoisers are mostly trained on synthesized data. The performance of CNN denoisers are highly dependent on whether the distributions of training noise and test noise are well matched.

Modern denoising methods often employ convolutional neural networks (CNNs) to learn the mapping function from noise to clean on a large collection of noisy/clean image pairs. These methods require accurate image model definitions, thus performance is limited in real-world cases. BM3D ) exploit a property of the noise or image structure to help denoising. (d) Our Noise2Blur needs some unpaired noise and clean images, as well as blurred counterparts of the noisy images to generate supervision.Ī large variety of denoising algorithms have been developed to deal with image noise. (c) Noise2Void uses just individual noisy images as training data. (b) Noise2Noise trains the network using pairs of independent noisy measurements of the same target. (a) Traditionally, the training of denoising network requires a large amount of paired noisy/clean images. Such complicated generation processes makes it difficult to estimate the latent noise model and recover the clean image from the noisy observation.įigure 1: Comparison of different training schemes. In real imaging systems, image noise comes from multiple sources, such as capturing instruments, data transmission media or subsequent post-processing. Noise often severely degrades the image, resulting in the loss of important image content. Noise is inherent to many imaging system, such as synthetic aperture radar images, undersampled magnetic resonance images and ultrasound images. Image denoising, which aims to restore a high-quality image from its degraded observation, is a fundamental problem in image processing. Experiments on several denoising tasks show that the denoising performance of N2B is close to that of other denoising CNNs trained with pre-collected paired data. These two networks are trained simultaneously and mutually aid each other to learn the mappings of noise to clean/blur. Using the new image pair, the denoising network learns to generate clean and high-quality images from noisy observations. Then, the noise map is added to a clean image to generate a new “noisy/clean” image pair.

First, the noise extraction network learns to output a noise map using the noise information from the denoising network under the guidence of the blurred labels. The N2B model consists of two parts: a denoising network and a noise extraction network. The training of the model requires only some (or even one) noisy images, some random unpaired clean images, and noise-free but blurred labels obtained by predefined filtering of the noisy images. We propose a new framework called Noise2Blur (N2B) for training robust image denoising models without pre-collected paired noisy/clean images.

Denoise online