Live broadcast picture pornographic detection and picture filtering method

Abstract

The invention discloses a live broadcast picture pornographic detection and picture filtering method, which comprises the following steps of: performing frame extraction and screenshot on a live broadcast stream, and numbering and storing the live broadcast stream in sequence; reading the screenshots with the numbers to detect the similarity of front and back pictures; after the similarity detection is passed, the pictures being sent to corresponding service providers for inspection; and fusing the detection results of the plurality of service providers. According to the live broadcast picture pornographic detection and picture filtering method, under the condition that the cost is not increased, the recognition results of multiple service providers are fused, and the recognition efficiency and accuracy are improved; and through a similarity filtering mechanism, the number of submitted pictures is greatly reduced on the premise of effectively guaranteeing the recognition effect, and the submission cost is saved.

Description

technical field [0001] The invention relates to a method for pornography detection and picture filtering of live pictures. Background technique [0002] With the advent of the Internet era, various live broadcast platforms emerge in endlessly. Due to national policy supervision and platform tonality requirements, live broadcast content needs to be legal and compliant, so pornographic and other violations are not allowed in live broadcast content. [0003] Against the above background, a system that can automatically identify, monitor and detect users' live broadcast content is needed. However, due to technical blind spots, and the live broadcast images are still, if the images are still or the changes are small, the cost of image detection will be high if all images are submitted for review at this time, due to the training models of various service providers. Different, the recognition effect of the training is also different, and it is impossible to comprehensively attack...

AI Technical Summary

Problems solved by technology

[0003] Against the above background, a system that can automatically identify, supervise and detect users' live broadcast content is needed. However, due to technical blind spots, and the live broadcast images are still, if the images are static or the changes are small, the cost of image detection will be high if all images are submitted for review at this time, due to the training models of various service providers. Different, the recognition effect of the training is also different, and it is impossible to comprehensively attack and detect pornographic content, so the effect of a single service provider is always limited, and the rate of false recognition and missed recognition is high

In addition, the service provider still has unstable factors, which may easily cause a single point of failure

Images