Improving Face Recognition by Integrating Decision Forest into GAN

Posture variation and self-occlusion are well-known factors that can compromise the accuracy and robustness of face recognition systems. There are a variety of ways to combat the challenges listed above, include using Generative Adversarial Networks (GANs). Nevertheless, many GAN methods cannot guarantee high-quality frontal-face images, which can improve recognition accuracy and verification when applied to multiple datasets. Recent results have proven that the two-pathway GAN (TP-GAN) method is superior to many traditional GAN deep learning methods that provide better face-texture details due to a unique architecture that enables the method to perceive global structure and local details in a supervised fashion. Although the TP-GAN overcomes some of the difficulties associated with generating photorealistic frontal views through the use of texture information provided by landmark detection and synthesis functions, it is difficult to replicate across different datasets. Particularly, under extreme pose scenarios, TP-GAN fails to further boost photo-realistic face frontalization image samples, minimizes the training time, and reduces computational resources, all of which result in substantially lower performance. This paper proposes simple adaptive strategies for overcoming TP-GAN’s inherent limitations. First, we incorporate the powerful discrimination capabilities of a decision forest into the discriminator of a TP-GAN. This method will result in a more stable discriminator model over time. Secondly, we acclimate a data augmentation technique along with a method which reduces training errors and accelerates the convergence of existing learning algorithms. Our proposed approaches are evaluated on three datasets, Multi-PIE, FEI and CAS-PEAL. We demonstrate both quantitatively and qualitatively that our proposed approaches can enhance TP-GAN performance by restoring identity information contaminated by variations in posture and self-occlusion, resulting in high quality visualizations and rank-1 individual face identification.