强化组合式生成对抗网络
|
摘要
近年来通过在生成对抗网络中加入条件标签,控制生成图像的类别或属性已经取得很大进展,但是生成图像的类别或属性的准确性有待提高。为此,在生成对抗网络的判别器中加入了强化学习,通过上一次的分类结果去指导当前的分类。另外为了让生成的细粒度图像更加逼真,使用注意力机制在只增加少量的计算损失下让图像有全局感受野,将多属性的星型生成对抗网络与自注意力生成对抗网络组合后的生成图像的质量较高。强化组合式生成对抗网络的最大均差达到0.036 93,最近邻指标效果较优,能自动化较准确地生成指定了某些属性的艺术图像,实验生成的图片也能用来解决缺乏数据的问题。
In recent years, great progress has been made in controlling the categories or attributes of generated images by adding condition tags to the generation adversarial networks. However, the accuracy of the category or attribute of generated image needs to be improved. In order to solve this problem,we add reinforcement learning to the generator of generation adversarial networks, which guides the current classification by previous. In addition, the attention mechanism is used which makes a global sensory field to the images with only a small amount of computational loss. We combines multi-attribute star generation adversarial networks with self-attention generation adversarial networks which improves the quality of generated. maximum mean discrepancy reaches to 0.036 93 and the 1-nearest neighbor classifier has a batter effect by reinforced-combined generative adversarial networks, which can generate the art images that certain attributes are assigned automatically and accurately. The generated images can also be used to address the lack of data.
In recent years, great progress has been made in controlling the categories or attributes of generated images by adding condition tags to the generation adversarial networks. However, the accuracy of the category or attribute of generated image needs to be improved. In order to solve this problem,we add reinforcement learning to the generator of generation adversarial networks, which guides the current classification by previous. In addition, the attention mechanism is used which makes a global sensory field to the images with only a small amount of computational loss. We combines multi-attribute star generation adversarial networks with self-attention generation adversarial networks which improves the quality of generated. maximum mean discrepancy reaches to 0.036 93 and the 1-nearest neighbor classifier has a batter effect by reinforced-combined generative adversarial networks, which can generate the art images that certain attributes are assigned automatically and accurately. The generated images can also be used to address the lack of data.
引文
[1] 汪旗, 贾兆红. 基于多示例学习的图像分类算法[J]. 计算机技术与发展, 2014, 24(4):88-91.
[2] 柯善武, 金聪. 基于优化空间金字塔匹配模型的图像分类[J]. 电子测量技术, 2017, 40(7):95-99.
[3] 韩辰希, 刘惠义, 商国中.视觉显著性纹理——色彩特征融合的图像目标分类[J]. 电子测量技术, 2017,40(11):94-98.
[4] KAELBLING L P, LITTMAN M L, MOORE A W. Reinforcement learning: A survey[J]. Journal of Artificial Intelligence Research, 1996, 4(1): 237-285.
[5] GUO L L, DING S F. Research progress on deep learning[J]. Computer Science, 2015, 42(5): 28-33.
[6] 张营营.生成对抗网络模型综述[J]. 电子设计工程, 2018, 26(5):34-37.
[7] WANG K F, GOU C, DUAN Y J.Generative adversarial networks: the state of the art and beyond[J]. Acta Automatica Sinica, 2017, 43(3): 321-332.
[8] 曹志义, 牛少彰, 张继威. 基于半监督学习生成对抗网络的人脸还原算法研究[J]. 电子与信息学报, 2018, 40(2):323-330.
[9] 孙玉红, 张元科, 孟静,等. 基于纹理和草图的图像铅笔画绘制[J]. 计算机应用, 2016, 36(7):1976-1980.
[10] 王坤峰, 鲁越, 王雨桐,等.平行图像:图像生成的一个新型理论框架[J]. 模式识别与人工智能, 2017, 30(7):577-587.
[11] 谢志峰, 杜胜, 郭雨辰,等.基于字典学习的HDR照片风格转移方法[J]. 图学学报, 2017, 38(5):706-714.
[12] SALLAB A E, ABDOU M, PEROT E,et al. Deep reinforcement learning framework for autonomous driving[J]. Electronic Imaging, 2017, 19(8): 70-76.
[13] 夏旻,宋稳柱,施必成,等.基于加权密集连接卷积网络的深度强化学习方法[J]. 计算机应用, 2018, 38(8):2141-2147.
[14] 王楠楠, 李洁, 高新波. 人脸画像合成研的综述与对比分析[J]. 模式识别与人工智能, 2018, 31(1):37-48.
[15] 林志洁, 丰明坤. 深度视觉特征与策略互补融合的图像质量评价[J]. 模式识别与人工智能, 2017, 30(8):682-691.
[2] 柯善武, 金聪. 基于优化空间金字塔匹配模型的图像分类[J]. 电子测量技术, 2017, 40(7):95-99.
[3] 韩辰希, 刘惠义, 商国中.视觉显著性纹理——色彩特征融合的图像目标分类[J]. 电子测量技术, 2017,40(11):94-98.
[4] KAELBLING L P, LITTMAN M L, MOORE A W. Reinforcement learning: A survey[J]. Journal of Artificial Intelligence Research, 1996, 4(1): 237-285.
[5] GUO L L, DING S F. Research progress on deep learning[J]. Computer Science, 2015, 42(5): 28-33.
[6] 张营营.生成对抗网络模型综述[J]. 电子设计工程, 2018, 26(5):34-37.
[7] WANG K F, GOU C, DUAN Y J.Generative adversarial networks: the state of the art and beyond[J]. Acta Automatica Sinica, 2017, 43(3): 321-332.
[8] 曹志义, 牛少彰, 张继威. 基于半监督学习生成对抗网络的人脸还原算法研究[J]. 电子与信息学报, 2018, 40(2):323-330.
[9] 孙玉红, 张元科, 孟静,等. 基于纹理和草图的图像铅笔画绘制[J]. 计算机应用, 2016, 36(7):1976-1980.
[10] 王坤峰, 鲁越, 王雨桐,等.平行图像:图像生成的一个新型理论框架[J]. 模式识别与人工智能, 2017, 30(7):577-587.
[11] 谢志峰, 杜胜, 郭雨辰,等.基于字典学习的HDR照片风格转移方法[J]. 图学学报, 2017, 38(5):706-714.
[12] SALLAB A E, ABDOU M, PEROT E,et al. Deep reinforcement learning framework for autonomous driving[J]. Electronic Imaging, 2017, 19(8): 70-76.
[13] 夏旻,宋稳柱,施必成,等.基于加权密集连接卷积网络的深度强化学习方法[J]. 计算机应用, 2018, 38(8):2141-2147.
[14] 王楠楠, 李洁, 高新波. 人脸画像合成研的综述与对比分析[J]. 模式识别与人工智能, 2018, 31(1):37-48.
[15] 林志洁, 丰明坤. 深度视觉特征与策略互补融合的图像质量评价[J]. 模式识别与人工智能, 2017, 30(8):682-691.