From October 22nd to 29th, the world's top computer vision experts gathered in Venice to participate in the ICCV 2017 International Computer Vision Conference to focus on the latest achievements in the field. The conference proceedings also represented the latest development direction and the highest level in the field of computer vision. The two ICCV papers submitted by Di Xinhan, an image algorithm engineer at NetEase Cloud Security (Eidun), were accepted by the conference and invited to participate in special seminars.
ICCV is the highest level conference in the field of computer vision. It is hosted by the IEEE and is held every two years worldwide. The ICCV paper acceptance rate has been very low, and it is the highest recognized level among the three international conferences on computer vision (the other two are CVPR and ECCV).
Di Xinhan at the ICCV site
At the special report of "CEFRL: Compact and Efficient Feature Representation and Learning in Computer Vision" of ICCV2017, Di Xinhan attended and participated in the discussion and development of the cutting-edge compact and effective features. And semi-supervised features. Invited guests at the seminar include well-known scientists in the field of deep learning AI such as Yoshua Bengio and Professor Pascal Fua.
Di Xinhan's two papers focus on the technical methods of improving the accuracy of semi-supervised classification and improving the quality and diversity of generated pictures. This method is applied to the image recognition of advertising, pornography, violent terror, and government affairs in the NetEase Cloud Security (Eidun) "content security" product line. It not only significantly improves the accuracy rate, but also collects less data in the training image. Next, it further improves the accuracy and recall of image classification. At the same time, using the GAN-Boost method, the number of difficult-to-collect images in the image training library is expanded and the image quality is improved, which reduces the cost of database collection for the product.
Di Xinhan's two selected papers presented at ICCV
In the paper, combining the information entropy and GAN network, to combat the input uncertainty of GAN network, two hypotheses are proposed under the theoretical framework of information entropy, and the principle of information entropy and noise channel is used to propose the Multiplicative noise channel The GAN-Boost training method proves that the two proposed methods reduce the loss of general GAN network information and improve the efficiency of GAN network using training data set information, and it is confirmed through experimental comparison that it improves the semi-supervised classification of GAN network images. The accuracy rate enhances the quality and diversity of images after training a small data set on a GAN network.
NetEase Cloud Security (Edun) is the first to launch the third generation of intelligent content security products in the industry. At present, it is already leading the industry in the areas of bad content identification such as smart identification, advertising filtering, violent terror and political identification. NetEase Cloud Security has continuously invested in research and development resources to improve the accuracy of intelligent identification, and has been widely used in well-known products such as NetEase Cloud Music, First Live, and OPPO App Store.