In recent years we have achieved a lot in trying to teach a computer to understand what he sees. However, with regard to the assessment of quality and aesthetic appeal, then we are faced with great difficulties. The difficulties that Google was able to overcome.
You can understand how the AI will understand whether the picture is of poor quality: artifacts, blurring, signs of compression – all this is to recognize not too difficult. But how to explain aesthetics? How to find the algorithm that will be able to understand what emotions is the picture, and is she beautiful?
The Google long and hard taught convolutional neural network on the examples of different images – for example, spectacular scenery. But this approach was not sufficiently difficult, because so images can be classified to just two types – good and bad. The new system, which was invented in the company is different and offers award each image ratings. This will allow for a more detailed evaluation of the quality of the picture.
How to achieve this from the computer? Enter NIMA – Neural evaluation of the image. This is a convolutional neural network that is trained to predict which picture we like, and we’ll give a high rating as an objective concerning quality and aesthetic appeal. NIMA exists thanks to successful neural networks responsible for recognizing objects. Thanks to them, the system understands that is the picture which in turn allows you to category root image on different factors.
New technology created by Google, will not only award pictures smart rating but also to ease the many difficult and time-consuming tasks like photo editing, optimization of quality and “fix” noticeable artifacts.
How NIMA will award picture ratings? To put it simply, there are two approaches. The first uses existing examples of “perfect” images, basing his conclusions on them. The second is “blind” and bases its findings on statistical models. Both approaches exists for one purpose: to evaluate the quality of the picture, which was correlated with the perception of the person. This helps to recognize the object in the picture, using existing databases like ImageNet.
Neural image rating
A typical assessment of the aesthetic appeal of the image allows the picture to award either a high or a low rating, but the neural evaluation of the image relies on a histogram of human ratings, and not just a binary system of evaluation. The histogram helps to understand the overall rating of the images and relates the views of different people. NIMA does not mention the picture either good or bad, and creates an approximate rating from 1 to 10. This system helps better predict a person’s reaction and more effective than any other.
In order to understand whether the technology you need to conduct a practical test. To do this, the researchers used photos, participating in contests. NIMA evaluated each picture according to their criteria and assigned them a rating. As it turned out, her assessment of beauty and quality was about the same as men. When this test was conducted several times, with successful results.
NIMA is also able to assess the quality in comparison. She correctly assessed the level of quality of one picture. The best was the photo without distortion, worst image with a great loss of quality due to compression.
Researchers are not only interested in the estimated abilities of the system, but in how she can apply that knowledge to make better images. If she knows what the picture is likely to appeal to someone, what prevents her to improve it? The answer is: nothing.
The following examples show how NIMA used their knowledge to make photos more attractive to humans. According to the algorithm, the picture gets better when modifying contrast. Also, the system itself is set up the brightness, saturation and shadows.
The development of NIMA gives to understand that this technology is available in many practical applications. For example, of the hundreds of images, the system can select the best. NIMA can be a helper, which in real time will comment on the potential. The system will be able to order a huge number of images in a short time. NIMA will become a competent (though not perfect) way to rate the pictures, predicting the reaction of the viewer.