基于残差结构的GAN网络的显著性预测研究  

Research on saliency prediction of GAN network based on residual structure

作　　者：赖远哲 陈向阳[1,2] 李旭东[1] 付星堡 曹倩倩 LAI Yuanzhe;CHEN Xiangyang;LI Xudong;FU Xinbao;CAO Qianqian(School of Computer Science and Engineering,Wuhan Institute of Technology,Wuhan 430205,Hubei,China;Hubei Key Laboratory of Intelligent Robot,Wuhan Institute of Technology,Wuhan 430205,Hubei,China)

机构地区：[1]武汉工程大学计算机科学与工程学院,湖北武汉430205 [2]武汉工程大学智能机器人湖北省重点实验室,湖北武汉430205

出　　处：《微电子学与计算机》2021年第8期95-100,共6页Microelectronics & Computer

基　　金：武汉工程大学教育创新基金(CX2019234)。

摘　　要：优化了简单生成对抗网络结构,用于更有效的通过对抗性实例训练得到视觉显着性图,减少假阳性产生和提高显著性.网络模型仍遵循传统生成对抗网络结构,第一阶段是由一个使用残差结构建的生成器组成,该模型的权值由显著图的下采样版本的二分类交叉熵损失(BCE)的反向传播计算得到的,训练得到更有效的显著图.预测结果由受训练的判别器网络进行生成阶段生成的显著图与真值图之间的二值分类处理.实验展示了改进生成对抗网络中的生成器的预测显著图的能力对整个网络性能提升,相较于其他显著图预测模型也有一定领先.The structure of simple generated countermeasure network is optimized,whichis used to obtain visual saliency map more effectively through antagonistic case training,so as to reduce false positive and improve saliency.The network model still follows the traditional generation countermeasure network structure.In the first stage,it is composed of a generator built with residual structure.The weight of the model is calculated by the back-propagation of the binary cross entropy loss(BCE)of the down sampling version of saliency map,and the more effective saliency map is obtained by training.The prediction results are classified by the trained discriminator network between saliency map and truth graph.The experiment shows that the improved ability to generate predictive saliency map of generators in the countermeasure network can improve the performance of the whole network,and it is also ahead of other saliency map prediction models.

关 键 词：视觉显著性 生成对抗网络 残差结构 交叉熵 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]