NWD-YOLOv5: A Mask Detection Algorithm for Complex Scenarios

Since the outbreak of covid-19 in early 2020, it has become a human consensus to effectively prevent infection with the covid-19 by wearing a mask. Therefore, it is significant to propose an object detection model that can accurately detect whether a face is wearing a mask or not in real time. However, in some complex public scenarios, some tiny target faces may be captured due to the distance of the camera from the target, which leads to great challenge of object detection. For most conventional object detection models, small targets are often missed. To solve the issue mentioned above, we make improvements to YOLOv5, which is called NWD-YOLOv5. First, to make our model much more focused on targets, we integrate the convolutional block attention model(CBAM) into YOLOv5. Then one more prediction head is additionally added to detect Tiny targets. At last, we replace the commonly used intersection over union(IoU) metric with the normalized gaussian wasserstein distance(NWD) metric to improve the accuracy of small target detection. After extensive experiments on mask dataset, the NWD-YOLOv5 has better performance than before. Based on the slightly improved AP results, the recall of the model is substantially improved. Compared to YOLOv5, NWD-YOLOv5 improves about 4.2% of recall. The link to the algorithm code is as follows: github.com/KartRiderZoey/nwdyolov5