Decryption out of thin air computer vision technology Sohu-月丘うさぎ

By , April 19, 2018 2:59 am

Decryption: "out of thin air" computer vision technology Sohu Lei Feng: CSDN net according to the original copyright belongs to the original "programmer" magazine, author Wei show ginseng, published in 2016 7. Lei Feng network (search for "Lei Feng" public concern number) has been authorized by the original author, for reprint please contact. Vision (Computer, CV) is a study of how to make the machine "see" science. 1963 Larry Roberts from MIT published the first doctoral dissertation in the field, "Machine Perception Three-Dimensional Solids", marking the beginning of CV as a new direction of artificial intelligence research. In the development of more than 50 years later, we have to talk about computer vision has recently let several "out of thin air" ability interesting attempt: super resolution reconstruction; image coloring; talk; portrait image restoration; automatic generation. As can be seen, these five layers of progressive layers, the degree of difficulty and interest has gradually improved. Due to the limited space, this article only talk about visual problems, do not mention too specific technical details, if you are interested in a part of the future, and then write a separate article discussion. Super-Resolution super resolution (Image) last summer, a "Waifu 2x" island applications in animation and computer graphics in a real fire. Waifu 2x with the depth of convolutional neural network (Convolutional Neural Network) technology, the image resolution can be improved by 2 times, but also to the image denoising. In simple terms, is to let the computer "out of thin air" and not some fill pixels in the original image, so that the cartoon looks more clearly true. We take a look at Figure 1, figure 2, really want to see childhood is so high dragon ah! Figure 1 "Dragon Ball" in super resolution reconstruction. The right side is the original, left for the super resolution reconstruction compared to Waifu 2x on the same frame super resolution reconstruction results in Figure 2 above Waifu 2x, low resolution and animation image noise, left for direct amplification results, right under the denoising and super resolution results for Waifu 2x however, the research of image super the resolution began around 2009, but due to the development of "deep learning", the Waifu 2x can make better effect. In the training of CNN, the input image for the original resolution, and the corresponding super resolution image as the goal, to make up the training of "image of" (Image Pair), after training the model can get super resolution reconstruction model. Waifu 2x depth network prototype based on the results of the Chinese University Hong Kong professor Tang Xiaoou team (see Figure 3). Interestingly, the study points out that the traditional.相关的主题文章:

Comments are closed