Computer Vision News - September 2018

Every month, Computer Vision News reviews a research paper from our field. This month we have chosen: Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising . We are indebted to the authors ( Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhan ), for providing us with great new images to illustrate our review. Their article is here . Introduction: Image restoration is a preliminary step in low-level computer vision with many applications. Although many very good approaches have been proposed, active research into better methods continues. The goal of image denoising is to recover clean image x from corresponding noisy image y, assuming y = x + v, v is commonly assumed to be additive white gaussian noise with standard deviation . Two special cases are single image super-resolution (SISR) and JPEG image deblocking. In this paper the authors investigated the construction of feed-forward denoising convolutional neural networks (DnCNNs) incorporating the progress in very deep architecture. DnCNN outperforms state-of-the-art methods handling both blind gaussian denoising (with unknown noise level), SISR and JPEG image deblocking. To give you a general idea of the capabilities of the DnCNN model for dealing with different denoising tasks, the Input Image below (a) is put together with 3 types of noise, each at two different noise levels. Gaussian noise with σ=15 (top left) and σ=25 (bottom left), low-resolution interpolation images with an upscaling factor of 2 (top center) and 3 (bottom center), and JPEG images with a quality factor 10 (top right) and 30 (bottom right). The white lines in the Input Image are only for us to differentiate the 6 different noise-type areas. The Output Residual Image (b) was normalized to the range [0,1] for illustrative purposes only. (c) is the denoised – restored Image. 4 Research: Beyond a Gaussian Denoiser Research by Assaf Spanier Computer Vision News DnCNN outperforms state-of-the-art methods handling both blind gaussian denoising (with unknown noise level), SISR and JPEG image deblocking.