基于改进Cycle-GAN的光流无监督估计方法CSCD

卷积神经网络为光流的计算提供了一种新的方式,但作为一种数据驱动技术,用于训练网络的大规模光流真值在现实世界中不易获取。为了解决这个弊端,基于Cycle-GAN的循环对抗机制,提出了一种光流无监督估计方法。首先,引入双判别器机制在生成器生成的光流样本的底层和高层特征上进行鉴别,迫使生成器提高光流生成的精度。其次,引入Spynet作为教师网络,在生成器训练前期对其进行指导,防止网络陷入模式崩塌。最后,改进损失函数,提出了光流一致性损失和轮廓一致性损失函数,进一步提升光流估计精度。实验结果表明,与现有的先进算法相比,提出的方法达到与有监督算法相同的精度水平。

Unsupervised Estimation of Optical Flow Based on Improved Cycle-GAN

Convolutional neural network provides a new perspective for optical flow estimation. However, as a data-driven technology, the GT value of optical flow used to train the network is inconvenient to obtain in the real world. In order to slove this problem, an unsupervised estima-tion method of optical flow is proposed based on the cycle adversarial mechanism of Cycle-GAN. Firstly, the dual discriminator mechanism (DDM) is designed to distinguish the optical flow generated by the generator on the low and high-level features, forcing the generator to improve the accuracy of optical flow generation. Secondly, Spynet is introduced as a teacher network to guide the generator in the early training stage and prevent the network from falling into model collapse. Finally, the loss function is improved. The optical flow consistency loss and the contour consistency loss function are proposed to further improve the accuracy of optical flow estimation. Experimental results show that the proposed method achieves the same level of accuracy as the supervised algorithm compared with the existing advanced algorithms.