Improved YOLOv3 Model for miniature camera detection

Abstract

The abuse of miniature cameras has severely violated information security and privacy. Several unsolved challenges such as small or multiple targets detection as well as complex background environments in terms of active laser detection still exist, resulting in difficulties in its practical application. In this paper, an active laser detection system is proposed to obtain high-intensity cat-eye reflection images. An improved YOLOv3 model, YOLOv3-4L, was introduced to detect the actual position of the target. In the YOLOv3-4L model, each image was resized to 608×608 to preserve image details. The scales of prediction were increased from three to four, and an additional feature map was used to extract more details. YOLOv3-4L exhibited excellent performance in detecting small targets. The experimental results show that the mean average precision achieved by YOLOv3-4L was 90.37%, compared to the 85.41%,87.81%, and 88.97% achieved by traditional YOLOv3, Faster R-CNN, and Single Shot Multi-box Detector respectively. The speed of the YOLOv3-4L model meets the requirements of real-time target detection.