摘要
针对目前视觉显著性预测网络模型训练时损失函数单一不能全面反映模型优劣,亦或网络模型非常复杂的情况,提出一种基于生成对抗网络(GAN)框架的自驱动显著性预测网络模型。其由两部分组成:生成器提取输入的原始图像特征生成显著性预测图;判别器用来分辨前一个部分生成的显著图和真实显著图的区分度。通过这种螺旋上升的对抗过程,期望能够生成与真实显著图相差无几的结果。实验表明,本网络模型在SALICON测试数据集上常规性能指标上可以获得不错的成绩,其中CC指标比对比方法提高近一个百分点。
At present,the saliency prediction network model is usually trained based on a single loss function,or the network model is very complex. In this paper,a self-driven saliency prediction network model based on the framework of Generating Countermeasure Network(GAN)is proposed. The network model consists of two parts:the generator is used to generate saliency prediction maps for the input original image,and the discriminator is used to distinguish the saliency maps generated by the former part from the real saliency maps. Through this spiral-up confrontation process,it is expected to generate saliency maps that resembles the ground truth. The network model can achieve good results on the conventional performance indicators of SALICOM test data set,and CC indicator can be improved by nearly one percentage point compared with the comparison methods.
作者
亢伉
KANG Kang(School of Computer Science,Baoji College of Arts and Sciences,Baoji 721016,China)
出处
《电子设计工程》
2020年第8期180-183,193,共5页
Electronic Design Engineering
基金
陕西省教育厅专项科学研究计划项目(19JK0036)
宝鸡市科学技术研究发展计划项目(2018JH-24)
宝鸡文理学院校级重点项目(ZK16009)。
关键词
视觉显著性预测
生成对抗网络
深度学习
显著图
visual saliency prediction
generative adversarial network
deep learning
saliency map
