An Improved Mobilenet-SSD Approach for Face Detection

Mobilenet-SSD is a lightweight network with high efficiency, which is widely used in the field of real-time face detection. Whereas, it fails to achieve similar high performances compared to region-based CNN methods. In this paper, we propose an improved Mobilenet-SSD approach by optimizing the feature map and the number of prior boxes of the original Mobilenet-SSD. These changes permit the proposed approach to get a high precision and recall in face detection. To assist further with face detection, the method of non-maximum suppression is employed to remove redundant candidate boxes. To evaluate the proposed method, we conduct experiments on the well-known FDDB benchmark dataset. For 300×300 input, the proposed method achieves 91.92% average precision (AP) at 39.0 frames per second (FPS) on GEFORCE GTX 1650. Throughout experimental results, we demonstrate that our approach achieves a considerable improvement on the AP with a slightly degraded speed compared with Mobilenet-SSD. In addition, our approach also outperforms Single Shot MultiBox Detector(SSD) in terms of speed and model size.