QCon, a global software development conference hosted by Geekbang Technology and InfoQ China, was held in Bohua Marriott Hotel, Shanghai from October 18 to 20, 2018.

In the special session of artificial intelligence and Deep learning practice (special session of solutions), Li Chaoguang, head of research and development of Qiniu Cloud deep Learning platform, was invited to attend and brought a wonderful speech with the theme of “Qiniu Deep Learning Platform supports hundred-million-level audit System”. Detailed explanation of content audit users encountered problems and optimal solutions, seven niuyun content audit system’s strong advantages and other content, for the guests to bring full of technical dry goods.



At present, the number of online video and live streaming users has exceeded 500 million and 400 million respectively, resulting in the production of a large number of illegal content and high management costs. Pictures and videos produced on the Internet may be carriers of undesirable information. Once spread, they will cause serious consequences and even endanger national security. With the government’s increasingly strict supervision of undesirable information, every website and platform is seriously treating and dealing with these risk factors, while the traditional way of human audit to deal with the new pictures and videos on the Internet every day has become an impossible task.

At the meeting, Li Zhaoguang believed that the challenges facing content security can be summarized into four aspects:

  • Audit efficiency is low. UGC platform with human audit mechanism takes 15-60 minutes to audit, which seriously affects user experience.

  • High human cost, audit personnel recruitment, hiring, training, site and other expenses bring a lot of expenses outside the main business, the cost is often up to one million;

  • The government implements the strictest content supervision standards. Once the content is violated, it will be punished immediately or even directly shut down. Negative public opinion will damage normal operation and corporate image.

  • The audit standards lag behind, the latest audit standards are difficult to be fully synchronized in the audit team in the first time, resulting in violations of the review of omissions, causing significant damage.

It can be said that content security has become the lifeline of each major platform, only under perfect management and audit, can prevent and avoid risks, avoid value damage.

In order to help users solve the problem of content review, Li Chaoguang said, “As a cloud service provider deeply rooted in the industry for many years, seven Niuyun has massive picture and video storage resources and perfect picture and video processing services. And, since the establishment of the seven NiuYun AI lab has created a unique depth of training platform, and the depth of training platform and depth of the whole business system is together, by the depth of the entire training platform to provide the basis of most of the training function, including iterative function, make the whole audit system iteration is very fast, the accuracy of the model will be high.”



It is understood that AVA, a container-based distributed learning platform, can efficiently schedule GPU resources, thus greatly improving model training efficiency and resource sharing rate. AVA opens up the process of data management, annotation, training, evaluation and release, and provides an efficient model development production line. AVA, a deep learning platform, has the following characteristics:

  • Easy to use: most functions are integrated into the Portal page, and training can be started by clicking the mouse;

  • Integrated labeling system: LabelX marking system, supporting video, image classification, detection, etc.

  • Cloud data set management: data support major cloud operators, easy data management on the cloud;

  • Convenient model management: one-click release, support fast iteration;

  • Flexible use of GPU resources: multi-machine and multi-card distributed training, multiple scheduling strategies;

  • Supports multiple frameworks, including MxNet, Pytorch, Caffe, and Tensorflow.

  • Flexible deployment: Supports public and private cloud deployment, and data does not leave the island.

  • Service adjustment: Service tasks are scheduled in batches and synchronously and asynchronously.



So, how does AVA, the deep learning platform of Qiuniuyun, help the auditing system? The Li Chaoguang said “seven NiuYun the contents of the latest security products” bright pupil 2.0 “using machine vision technology, artificial intelligence to pornography, violence will massive video image data, and identification of the sensitive character, to ensure the Internet, radio and television, new media industries, as well as government agencies such as health, the content of the data transmission for seven NiuYun ecological security products, To this end, Seven Niuyun launched a one-stop smart phone review + commissioner review service system, which can filter 99.5% of violations of the content, helping users to save 80% of the cost.”

Under these management, Qiuniuyun can provide customers with three zero services, namely zero risk operation, zero threshold use, zero burden management, users only need to open qiuniuyun button can use.

In terms of industry landing, Qiniuyun has been in the forefront of The Times in the field of artificial intelligence for a long time. “Pupil 2.0” can be widely used in video platform, live broadcast platform, e-commerce platform, forum community, social media, public network disk, radio and TELEVISION and other fields. Currently, it provides services for customers of various business types, such as Ping An of China, Faceu, Blued, Mtime, Tangdou and so on.

Seven NiuYun

As a leading enterprise in the industry, Seven Niuyun started with cloud storage and has been deeply engaged in the industry for seven years. It always takes “shortening the distance from idea to product” as its original intention and mission. Relying on its strong technical advantages and good reputation, it manages more than 200 billion files, adds more than 2 billion files every day, and accesses hundreds of billions of data every day. It serves more than 700,000 corporate customers and indirectly serves more than 80% of the country’s netizens. In the future, Qiuniuyun will be committed to providing sustainable intelligent video cloud services for all industries and applications, helping enterprises to quickly access the cloud and creating broader business value.