Computer Vision News - September 2018

(a) find rotated derivative filters in different directions; (b) contain first, second and higher-order derivative filters; (c) adaptive diffusion learned through the nonlinear functions. Since the revolution in deep learning, several methods have been proposed to overcome the limitations of prior-based approaches, by eliminating iterative optimization at the testing stage. One of them is VDSR a neural network architecture whose purpose is single image super resolution (SISR). The VDSR network learns a mapping between low resolution images and their respective high resolution images, learning to predict their residual image (noise). However, unlike DnCNN, this network only deals with and was only tested on SISR tasks, and doesn’t use newer techniques such as batch normalization. Method: The DnCNN network the authors developed has a simple basic structure made up of three types of layers: (1) Conv+ReLU, (2) Conv+BN+ReLU, and (3) Conv, shown in the figure below in yellow, blue and orange, respectively. Research 6 Research Computer Vision News Training details Network name For gaussian denoising with known noise level: the authors considered three noise levels: = 15, 25 and 50. Patch size was 40x40 Network depth (number of blue layers): 17 DnCNN-S. train the DnCNN model for a range of the noise levels ( ) as 0 through 55, and the patch size as 50X50. DnCNN-B B is for blind (noise level is unknown) Trained on color version of the BSD68 dataset CDnCNN-B B is for blind / C is for color Trained simultaneously on 3 image denoising tasks: blind gaussian denoising, SISR, and JPEG deblocking. DnCNN-3