Abstract: For object detection, one immediate problem is the insufficiency of feature extraction on small objects, which is easy to make false detection and miss the inspection on small targets. To solve the problem of small object detection, an improved detection algorithm based on YOLOv5 was proposed. The algorithm uses the method of Mosaic-8 on data augmentation. A shallow feature map is added to the YOLOv5 network and loss function is adjusted to improve the sensibility of network on small targets. The target box regression formula is modified to solve the problem of gradient disappearance in training process, which realized accurate precision on small targets. The improved algorithm is applied to mask wearing detection under crowed environment. Experimental results show that the proposal algorithm has stronger feature extraction ability and higher detection accuracy on small object detection compared to the original YOLOv5 algorithm.