Conventional face recognition (FR) systems face challenges with varying poses, scales, and occlusions, particularly in dense environments where interpersonal occlusion is common. Existing methods using rectangular bounding boxes (BBs) often result in inaccurate detections and lower FR accuracy, particularly when landmark-based alignment fails. To address this, we propose a novel approach integrating Bound YOLO-v7 with a context module to improve face boundary marking, extend the receptive field, and preserve facial contours. Supported by a newly annotated boundary dataset, the method fills the gap in high-quality benchmark data for facial boundary segmentation. In the offline phase, Bound YOLO-v7 extracts face contours, while in the online phase, FaceNet identifies multiple faces in real time. The proposed method achieves a detection rate of 99.83% with mAP values of 0.995 and 0.979 for mAP@0.5 and mAP@0.5:0.95, respectively, and a confidence score of 0.42 ms at 41.3 ms. The inclusion of the context module results in mAP@0.5 scores of 99.5% (no occlusion), 96.0% (slight occlusion), and 89.0% (severe occlusion). This approach outperforms the existing method and balances detection accuracy and computational efficiency.