New research from Google shows that in the highly subjective field of photography, machines can be just as aesthetically pleasing as humans, producing images that photographers like. To see how it works, read on!

Machine learning (ML) has performed well in many well-targeted areas. Tasks with a clear distinction between right and wrong answers will aid in training and allow the algorithm to achieve preset goals, such as accurately identifying objects from images or reasonably translating language. However, there are also many areas of task that are difficult to evaluate objectively, and when it comes to subjective problems such as evaluating the beauty of a photograph, which has to do with each person’s aesthetic, machine learning is not a good fit.

Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include information robot, editing robot, writing robot and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.

Photo: A professional photograph of Jasper National Park

To study how machine learning learns subjective concepts, Google has introduced an experimental deep learning system for artistic creations. The system mimics the work of a professional photographer. Here’s how it works: Browse through landscape images from Google Street View, analyze the best composition, and then perform various post-processing to create a pleasing image.

Through 40,000 panoramas of the Alps, banff and jasper national parks in Canada, BigSur and Yellowstone national park in California, the virtual photographer created a number of impressive images, some of which were even professional, said the professional photographer.

Although the aesthetic feeling in the photos of the training model can be simulated by using data sets similar to those in AVA system, the direct processing of the photos by AVA system may have some defects in aesthetic feeling, such as oversaturation of the photos. Again, the tagging data set required to properly learn beauty from multiple aspects through supervised learning is difficult to gather, so this is not a good approach.

Their method requires only a few high-quality photos and does not require before-and-after image comparisons or additional tagging. The system automatically breaks down beauty in photographs into different aspects, each of which can be learned separately from negative examples of opposite image manipulation.

By semi-orthogonalizing image processing, quick and independent optimization steps can be found to beautify images in terms of composition, saturation /HDR levels, and light and shade tension:

Figure: Figure (a) is a panorama, figure (b) is cropped, Figure (c) is saturation and HDR optimization of Figure (b), and Figure (d) is the effect after applying dramatic tension mask.

They used traditional image filters to generate negative examples including saturation, HDR detail and composition, and also introduced a special operation called dramatic tension mask, which is produced in conjunction with learning the concept of light and shade tension.

The negative examples were generated by applying a set of image filters to randomly adjust the brightness of high-quality photos, making them worse. In training, they used generative adversarial networks (GAN), in which the generative network creates a mask to improve the light in the negative example, and the discriminant network tries to distinguish the improved light from the sample photo.

Unlike shape-fixedfilters such as Vignette, dramatic tension masks add a content-aware brightness adjustment section. The natural competitiveness of GAN training greatly enriches the ability to adjust image characteristics, and more training details can be seen in the paper.

Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.

No one would dispute that attention is one of the scarcest, most valuable and most rationally utilized resources in this age of information fragmentation. Relying on the independently developed Giiso engine, the Zhisou team created the first intelligent media platform Tianzexun APP, which can intelligently answer users’ various relevant information according to various commands or text interaction commands. And can be based on the user’s personalized use characteristics and continuous learning, continuous tracking of users interested in the unique content. At present, the day smart news APP6.0 version has been updated iteration, can be used to download the application market.

Above, Chen Ruchu’s humble opinion! I’m sorry if I offended you.

For more information about ai robots, please visit Giiso Zhisou: http://www.giiso.com/. Thank you