An optimized deep learning framework for temporally stable real-time face detection in unconstrained video streams

Face detection in real-time video streams remains challenging because of dynamic variations in illumination, pose, motion blur, and partial occlusion. This paper proposes an optimized deep learning framework that integrates YOLOv12-based face localization, capsule-based pose-aware region-of-interest representation, Deep Q-Network-guided temporal decision stabilization, and Harris Hawks Optimization for accuracy–stability–latency parameter tuning. Experiments conducted on the Selfies & Video Dataset using subject/video-level cross-validation achieved an accuracy of 94.09%, precision of 94.77%, recall of 89.47%, F1-score of 90.84%, and Matthews Correlation Coefficient of 0.9334. Compared with YOLO-FaceV2, the proposed framework improved accuracy by 0.49 percentage points, precision by 0.57 percentage points, and MCC by 0.54 percentage points. The proposed temporal decision mechanism also reduced detection flicker and abrupt bounding-box variations across consecutive frames compared with frame-wise detector variants. With an inference speed of 31.4 frames per second and an average latency of 31.8 ms per frame, the framework provides an effective balance among detection accuracy, temporal consistency, and real-time computational efficiency for unconstrained video applications.