Google super resolution RAISR: fuzzy picture becomes clear, the operation speed is ten times faster – Sohu technology every day millions of images are shared on the network, storage, users to explore the world, research topics of interest, or share with friends and family vacation photos. The problem is that a large number of pictures to be limited by the pixels of the camera equipment, in order to be in the cell phone, tablet or network restrictions are artificially compressed, reducing the quality of the picture. Now the high resolution display screen is popular, family and so on mobile devices, the low resolution picture into a high-definition version, and can view and share on a variety of devices, is becoming a huge demand. Recently, Google launched a new RAISR technology, its name is "Rapid and Accurate Image Super-Resolution", meaning "super resolution technology fast and accurate". RAISR this technology can use machine learning, the low resolution images into high resolution images. It can achieve or even exceed the current super-resolution solution, while the speed increase of about 10 to 100 times, and can run on ordinary mobile devices. Moreover, Google technology can avoid aliasing (aliasing artifacts). Prior to the adoption of the up and down sampling method, the low resolution image reconstruction for larger size, more pixels, higher picture quality technology. The best known method of sampling is a linear method, which is to add a new pixel value by a simple, fixed combination of known pixel values. This method is very fast because of the use of a fixed linear filter (a constant convolution for the whole picture of the non discriminatory processing). But it has little to do with the vivid details of the hd. As the following picture, the picture is very blurred sampling, it is difficult to be called quality improvement. The left is the original picture; right up sampling after picture. For RAISR, Google another way to use machine learning, with low resolution and high resolution images of the training program, in order to find out can be selectively applied to each pixel of low resolution picture of the filter, thus generating a comparable to the original image details. There are two methods for training RAISR: either way, RAISR filters are based on edge features of training images: brightness and color gradient, plain and texture etc.. This is also affected by the direction (direction, edge angle), intensity (strength, sharper edge strength) and viscosity (coherence, a measure of edge orientation). The following is a set of RAISR filters, learned from ten thousand pairs of high and low resolution images (low resolution images are sampled). The training process takes about 1 hours. Note: 3 times super resolution learning, obtained by 11× 1相关的主题文章: