Multi-view Face Detection Using Hybrid Generative Adversarial Networks and Multi-Task Cascaded Convolutional Neural Network

Abstract

Multi-view face detection is the fundamental technology of an open human-computer interaction environment. Despite satisfactory development in frontal face detection, non-frontal extreme poses are still persistent challenge. In real-world scenarios, non-frontal views are common due to camera positioning, subject movement, or occlusion. As the head rotates, essential facial features such as the eyes, nose, and mouth become partially occluded, significantly degrading detection performance. To address the issues, we employed a two-stage solution with the combination of face frontalization and detection, Initially, Couple agent pose guided Generative Adversarial Network employed to generate a good quality frontal face images from their profile counterparts, thereby enhancing the effectiveness of conventional detection methods. This followed by mitigate application of Multitask cascaded convolutional neural networks for detection. By integrating CAPG-GAN for frontalization and MTCNN for detection, our hybrid solution framework improved the ness and face detection across diverse poses and occlusion conditions. The performance of the suggested hybrid CAPG GAN-MTCNN model was evaluated on two benchmark datasets: MultiPIE and CAS-PEAL-R1. The proposed model attained high structural similarity in generated frontal images with SSIM measures of 94.5% and 91.45% for Multi-PIE and CAS-PEAL-R1, respectively. Additionally, Human perceptual experiments also confirmed the visual quality and identity consistency of frontalized faces with near-perfect ratings even for extreme profiles. Based on our standard face detection metrics, our proposed CAPG- GAN-MTCNN model significantly improves multi-view face detection performance under extreme pose variations and occlusions. On the Multi-PIE dataset, our proposed approach achieved an average IOU of 0.792 to 0.949 and Recall of 0.885 to 0.976, surpassing the baseline MTCNN advanced and also on the CAS-PEALR1 dataset our proposed model achieved an average IOU of 0.756 to 0.935 and Recall of 0.86 to 0.95, surpassing the baseline MTCNN++ . The PR curves are also proof of the performance of the proposed mode, on the Multi-PIE dataset, the model accomplished an AUC of 0.985, whereas 0.965 on the CAS-PEAL-R1 dataset. These all results confirm that the CAPG-GAN-MTCNN framework offers a substantial advantage in detecting faces under extreme up to 90?? pose challenging, making it a strong candidate for real-world face detection applications.