基于条件生成对抗网络的视频显著性目标检测

Salient object detection based on conditional generative adversarial networks for video

针对传统的显著性检测方法存在着流程复杂,计算成本高,特征学习不足等问题,受生成对抗网络以及弹性网络的启发,提出一种基于条件生成对抗网络(cGAN)与L1,L2范式联合正则的视频显著性目标检测方法。方法需训练2个模型:生成器和判别器。生成器尽可能生成与真实值一样的显著图来迷惑判别器,使其难以辨别预测的显著图的真实性。判别器则尽可能准确地区分“假”显著图。实验表明:所提方法在两个公开视频数据集上的检测效果都超过了当前主流方法,且算法流程简单,运算效率更高。

Aiming at the problems such as complicated flow,high computational cost and insufficient feature learning that traditional saliency detection methods have,inspired by generative adversarial networks and Elastic networks,a video salient object detection method based on conditional generative adversarial networks(cGAN)combined with L1 and L2 paradigms is proposed.The proposed method requires to train two models:generator and discriminator.As much as possible,the generator generates the same saliency map as the truth value to confuse the discriminator,making it difficult to discern the authenticity of the predicted saliency map.The discriminator,on the other hand, distinguishes the“false”saliency map as accurately as possible.Experiments show that detection effect of the method outperforms the current mainstream methods in both open video datasets, and the algorithm has simple flow and higher efficiency.