Abstract
Facial recognition performance often degrades significantly when processing profile views compared to frontal views. This paper presents a Dual-discriminator Generative Adversarial Network (GAN) designed to synthesize frontal face images from profile photos. By utilizing a fixed, pre-trained encoder and two specialized discriminators—one for identity consistency and one for visual authenticity—the proposed method generates synthetic frontal views that maintain the subject's unique features. Experimental results on the Bosphorus dataset demonstrate that the generated images can successfully fool state-of-the-art recognition models like FaceNet.
Introduction
While modern facial recognition systems achieve near-perfect accuracy on frontal images, profile views remain a significant challenge. The feature representations, or eigenvectors, extracted from side-view images differ substantially from those of frontal views, even for the same individual. Traditional approaches, such as 3D reconstruction or massive multi-view data collection, are often computationally expensive or brittle. This research proposes a generative solution: transforming profile images into frontal views to leverage existing high-performance frontal recognition models.
Methodology
The core of the proposed system is a GAN architecture that balances identity preservation with image realism.
System Architecture and Training Flow
The framework employs a single generator and two distinct discriminators to provide specialized feedback during training.
To ensure stable training, the authors implemented an artificial intervention loop. This logic monitors the loss levels of the generator and discriminators, dynamically selecting which component to train to prevent the discriminators from overwhelming the generator and causing vanishing gradients.
The Generator: Encoder-Decoder
The generator is structured as an encoder-decoder network.
A key design choice is the use of a fixed, pre-trained encoder. By freezing the encoder (trained on a separate recognition task), the authors ensure it consistently extracts meaningful identity features. The DCGAN-based decoder is then trained to "translate" these features into a frontal face image.
Dual Discriminators
- Identity Discriminator: This network compares the input profile with the generated frontal face to ensure they belong to the same person.
- Authenticity Discriminator: This network focuses purely on visual realism. It utilizes a PacGAN structure, which analyzes multiple images simultaneously to prevent "model collapse," a common issue where the generator produces a limited variety of images.
Experiments and Results
The model was evaluated using the Bosphorus dataset, comprising 4,666 2D images of 105 individuals.
Identity Preservation
The dual-discriminator approach successfully bridged the "language gap" between the encoder and decoder. By avoiding triplet loss during encoder pre-training—which can discard common facial characteristics—the system retained enough information for the decoder to reconstruct believable faces.
Recognition Accuracy
To validate the utility of the synthetic images, the authors used FaceNet to classify the results. Notably, 16.4% of the synthetic images successfully fooled the classifier into identifying them as genuine frontal photos of the target individual. Furthermore, the training strategy effectively maintained diversity in the generated outputs, avoiding model collapse.
Conclusion and Future Work
The Dual-discriminator GAN provides a robust framework for profile-to-frontal face synthesis. While the generated images still exhibit some unnatural characteristics detectable by deep networks, they represent a significant step toward making profile face recognition more practical. Future research will focus on more complex generator architectures and refining the "hallucination" of features occluded in profile views.
Summary updated