社交媒体中图像歧视情感识别及去除关键技术的研究

Social images and their tags and discussions in texts may intentionally or unintentionally reveal and spread discrimination to others or individual groups in the sharing process. Based on technology of image recognition and characteristic of social media, this project proposes to carry out discrimination analysis and discrimination-free method of social media. The proposed project contains the following innovative researches. First, a deep mixture algorithm will be developed which will be capable of recognize large numbers of discrimination-sensitive object classes and attributes accurately and predicting types of image discrimination effectively. Second, a-GAN (Generative Adversarial Network) algorithm will be developed to automatically generate perceptually-similar and discrimination-free image patches to replace arbitrary-sized discriminatory image components for preventing direct image discriminations.The research achievement will provide technology support to better social media sharing. Moreover, it have potential to be applied to other domains.

社交媒体中图像共享可能由于有意无意的通过图像、标签和文本的讨论对个人或者群体传播着歧视的态度。本项目在图像识别及社交媒体特点基础上，开展社交媒体歧视态度分析及去歧视的方法，包括（1）提出一种深度混合算法识别大规模歧视敏感的对象类和属性，并能够有效预测图像歧视情感类型；（2）提出一种基于生成对抗网络的算法，自动生成视觉相似并无歧视的图像块，实现歧视敏感对象类的替换。取得的成果将为社交网络的健康发展提供技术支撑，同时还可以应用在更多的领域。

社交媒体中图像共享可能由于有意无意的通过图像、标签和文本的讨论对个人或者群体传播着歧视的态度。本项目在图像识别及社交媒体特点基础上，开展社交媒体歧视态度分析及去歧视方法的研究，从歧视关系检测、去歧视生成图像两个方面提出了解决方案，包括（1）提出一种基于关系模型的歧视敏感对象的检测方法，有效预测图像歧视情感类型；（2）提出一种基于生成对抗网络的算法，自动生成视觉相似并无歧视的图像块，实现歧视敏感对象类的替换。取得的成果将为社交网络的健康发展提供技术支撑，同时还可以应用在更多的领域。