The market capacity of picture identification service is huge. As the most popular entrepreneurial field in the mobile Internet industry, mobile social APP produces a large number of pictures every day, and there are numerous pornographic pictures mixed among them. Therefore, it is a very difficult task to identify and eliminate pornographic information efficiently and accurately.
In addition, the popularity of mobile live streaming has also led to a great increase in the demand for pornographic image verification. Especially for small and medium-sized development teams, live streaming platforms are likely to face risks in pornographic verification due to human supervision issues. The independent research and development of yellow detection function or the increase of audit personnel will increase the expenditure of products and services, causing additional pressure to the early development. Using artificial intelligence image recognition technology to carry out efficient and accurate automated yellow detection service can reduce the technical threshold for enterprises to use yellow detection service, and help enterprises effectively reduce the investment of related labor costs.
How to define sexiness and pornography
△ Conventional neural networks and deep neural networks
Machine learning is the core of artificial intelligence. To put it simply, it is: the use of a set of general algorithms — generic algorithms, the establishment of data logic, the use of the mechanism of imitating human brain to explain the data, so that the machine automatically learn good features, so as to reduce the process of manual audit.
To teach a machine to recognise pornographic images, for example, it would have to “train” it with thousands of sample images to pick up the features of pornographic images and keep them in its memory. Any point in each image contains a value of brightness, hue and saturation. By setting the range of these three values, the machine can identify “flesh colour” and guess which areas of skin are exposed in the image.
The most obvious feature of pornographic pictures is the large proportion of human skin color in the picture. When the machine identifies areas similar to human skin color in the picture, it needs to further confirm the source of the areas to see whether they are the heroine without clothes or normal objects. Assume that two pieces of yellow area or two arms, two legs are another area is to the body, these areas the proportion of length, width values accord with human body size, and position meet certain geometric relationships to one another, though, there was a big possibly pornographic images, if the area between the size and position, not like a person’s body can exclude suspected of pornographic images.
△ Calculates the geometric relationship between skin color regions
△ Picture discrimination standards
Pornography: Exposure to sensitive parts of the body, including explicit shots, images depicting sexual acts and pornographic scenes.
Sexy: Scantily clad but without exposing sensitive areas.
Normal: Not pornographic, not sexy pictures.
The identification criteria of pornography and art are determined by human beings. Theoretically, machines can be made to meet their own standards by means of deliberate training and adjustment of thresholds. The more pornographic pictures there are, the more diverse the styles and scenes are, and the more accurate the machine learning results will be.
A major advantage of machine learning is that it can use big data samples to continuously improve the identification accuracy during the learning process. Thanks to the improvement of computer speed, the rise of large-scale clustering technology, the application of GPU and the emergence of numerous optimization algorithms in recent years, the training process that takes several months can be shortened to a few days or even hours, and machine learning can be widely used to greatly improve the efficiency of yellow discrimination.
Artificial intelligence image yellow: machine learning and human review
△ and beat cloud intelligent discriminating yellow workflow
Youpai Cloud’s “intelligent yellow discrimination” function will automatically identify live broadcasts, videos, pictures and other content. At present, the complete process of a picture yellow identification is to take it to the yellow identification center, after completion, and then send the results to the picture review platform for final confirmation. For suspected pornographic pictures, they will be reviewed and confirmed by human beings, and this part will be continuously reduced with the increase of training times, so as to help the operation team save the cost of human review.
How to carry on the live broadcast of yellow
Under normal circumstances, live video porn identification services use video screenshots, image recognition, voice review, barrage monitoring, keyword extraction and other ways to identify pornographic content.
The verification of live video can follow the following steps: identify whether there are physical signs of people in the image and count the number of people; Recognize the gender and age range of the figure in the image; To identify the skin color of the character and the extent to which the body organs are exposed; Recognize the figure’s body outline, analyze the action behavior; Key words of audio information are extracted to judge whether there is sensitive information; Analyze the text content of the barrage screen in real time to judge whether there is any violation in the current video. The frequency of video key frames per minute can be set independently by the customer, ranging from 1 second to tens of seconds. For example, the key frames can be collected once every 5 seconds by default for identification.
Computer is too easy. I’m a shooter SKR, SKR, SKR, SKR
The AI was so good that it even took my job