基于双通道循环一致性GAN的跨视角步态识别研究

步态识别系统在生物识别领域显示出巨大的潜力,然而步态识别的准确性很容易受到视角的影响。为解决这一问题,提出的方法基于循环生成对抗网络(cycle generative adversarial network, Cycle-GAN)的网络结构,结合新的特征提取模块以及多重损失函数,提出了一种用于跨视角步态识别的网络模型双通道循环一致性生成对抗网络(two-channel cycle consistency generative adversarial network, TCC-GAN)。该网络首先将步态能量图像从任意视角转换为目标视角图像,然后进行比对从而实现身份识别。TCC-GAN分别引入了重建损失、视角分类和身份保持损失来指导生成器生成具有目标视角的步态图像并同时保留身份信息。为了避免可能存在的模式崩塌问题,并保证各个输入和输出以有意义的方式进行映射,模型中还利用了循环一致性损失。数据集CASIA-B和OU-MVLP上的实验结果表明:所提TCC-GAN模型的跨视角识别率高于目前大多数其他基于GAN的跨视角步态识别模型。

Cross-view gait recognition based on two-channel cycle consistency generative adversarial network

Gait recognition systems have shown great potentials in the field of biometric recognition. However, the accuracy of gait recognition is easily affected by a large view angle. In order to solve this problem, this paper proposed an improved model two-channel cycle consistency generative adversarial network(TCC-GAN). The network could transform the gait energy image from arbitrary views to the target view and then perform identity recognition. Besides reconstruction loss, it introduced view classification and identity preserving loss to guide the generator to produce gait images of the target view and keep identity information simultaneously. In order to avoid the model collapse problem that might exist and ensure that each input and output were mapped in a meaningful way, this paper also utilized the recently proposed cycle consistency loss, which could facilitate that the target image and the source image under the target view had the same feature distribution as possible. The experimental results on the datasets CASIA-B and OU-MVLP indicate that the proposed TCC-GAN model can obtain higher accuracy than other state-of-the-art GAN-based cross-view gait recognition models.