GCB-YOLO7: Enhancing Lightweight Face Detection with Ghost Module and Convolutional Block Attention Module

Abstract

Convolutional Neural Networks (CNNs) have significantly advanced object detection recently. However, achieving the right balance between model size and detection accuracy is a persistent challenge. While large models provide high accuracy, their computational demands make them impractical for real-time or embedded applications. Conversely, lightweight models are more efficient but often sacrifice accuracy. To address this, we proposed GCB-YOLO, a lightweight face detection model that balances compactness and performance. The design incorporates the Ghost Module (GM) to reduce model size, the Convolutional Block Attention Module (CBAM) to enhance feature representation, and the SiLU activation function to improve detection accuracy. Our evaluation on the WIDER FACE dataset demonstrates the effectiveness of GCB-YOLO. On the medium subset, it achieved 90.0% precision, 79.6% mAP50, and an 81.04% F1 score, surpassing other state-of-the-art lightweight models. The model maintained strong performance on the challenging hard subset with 79.3% precision, 60.5% mAP50, and a 65.37% F1 score. For the efficiency metrics, the parameter count was reduced by 49.44% (from 6.015M to 3.041M), and CPU inference speed improved from 4.265 FPS to 5.130 FPS. These results highlight GCB-YOLO’s suitability for real-time face detection on resource-constrained devices.