An Efficient Face Detection and Gender Classification Approach Integrating the Speed of YOLOv9 with the Accuracy of ResNet50

Artificial Intelligence (AI), particularly deep learning models, plays a pivotal role in tasks such as face detection and gender classification. This study aims to enhance the performance of computer vision systems by developing a hybrid model that integrates YOLOv9 with a modified version of ResNet50, addressing the common trade-off between accuracy and inference speed found in traditional approaches. To achieve this, a custom dataset was collected from real-world conditions within a university campus, testing the performance of multiple You Only Look Once (YOLO) models and Convolutional Neural Network (CNN) architectures. Experimental results revealed that YOLOv9 achieved the highest inference speed of 332 ms/image at 3.00 Frames Per Second (FPS), while ResNet50 demonstrated superior accuracy in gender classification as a two-stage detection model, albeit with slower performance. To resolve this trade-off, ResNet50 was modified for both speed and accuracy, and then structurally embedded into the YOLOv9 framework. Specifically, the Cross Stage Partial Network (CSPNet) and Efficient Layer Aggregation Network (ELAN) layers of YOLOv9 were replaced with modified ResNet50 feature extractors, while the Global Local Attention Network (GLAN) layer was retained to preserve effective feature fusion. This integration significantly improved both facial and object detection performance. The proposed hybrid model outperformed individual models, achieving a peak mean Average Precision (mAP) of 97.2%, with 97% precision, 93.4% recall, and an inference speed of 103.89 ms/image (9.62 FPS). These results demonstrate that the proposed model effectively balances accuracy and speed, making it highly suitable for real-time applications such as smart surveillance and security systems.