GPU ACCELERATED BIAS MITIGATION IN FACE DETECTION USING VARIATIONAL AUTO ENCODERS AND ONNX RUNTIME

Abstract

Bias in automated face detection—where certain underrepresented groups consistently experience degraded or erroneous inference—remains a significant challenge in computer vision. This thesis proposes an efficient, fairness-aware face detection framework that mitigates bias across both demographic groups and other facial attributes, while leveraging GPU acceleration for real-time performance. The core contribution is a novel joint training approach that integrates a Variational Autoencoder (VAE) branch and a reinforcement learning–based rarity-driven resampling strategy into a pre-trained Faster R-CNN detector. This design enables the model to learn latent representations of underrepresented facial features (e.g., rare demographic or appearance traits) and rebalance the training process accordingly—notably without requiring explicit demographic or attribute labels. A rarity-aware sampling policy, guided by the VAE’s latent space and treated as a black-box optimization problem via reinforcement learning, preferentially up-samples infrequent examples during training to improve fairness for minority groups. Using face detection with a ResNet-50 backbone as a challenging testbed, we demonstrate that this approach consistently narrows the performance gap between well-represented and underrepresented groups without degrading overall accuracy. Trained on the CelebA dataset and evaluated on CelebA and FairFace benchmarks, the proposed model achieves significantly higher detection precision (mAP) and localization quality (mean IoU) compared to a baseline Faster R-CNN detector without VAE integration. Specifically, on CelebA, we observed improvement in mean IoU for all attributes (40 attribute groups) ranging from a minimum of +0.8% (Blurry, N = 1926, p < 0.05) to a maximum of +5.3% (Rosy Cheeks, N = 3206, p < 0.001). Similarly, on the FairFace dataset (14 race and gender subgroups), mean IoU improved by +1.3% (Female, N = 30001, p < 0.001) up to +3.8% (Male, N = 31971, p < 0.001). To ensure practical deployment, the trained model was exported via ONNX Runtime and optimized for GPU inference. On GPU, the ONNX-exported model attains nearly a 19× throughput speed-up over CPU execution, enabling real-time face detection. In summary, this research demonstrates a fast and fair face detection method that significantly improves performance on potentially biased datasets without sacrificing efficiency or accuracy.