NWD-YOLOv5: A YOLOv5 Model for Small Target Detection Based on NWD Loss

Target detection technology is vital in modern social management and urban security, with small target detection— accurately identifying tiny objects in large scenes—posing a key challenge for improving surveillance system accuracy and real-time performance. The YOLO series is still the mainstream target detection model, for the YOLOv5 model, as the IoU metric it adopts still uses the geometrical overlap of the detection frame as the basis of the loss calculation, while the experiments have proved that this metric is unable to effectively utilize the small-scale characteristics of small targets. Therefore, improving the loss metric in the existing framework can help the model to pay more attention to the spatial semantic features of small targets in the image in the training phase of the network, respectively, and thus improve the detection effect. In this paper, we propose the NWD-YOLOv5 model, i.e., we introduce the NWD loss that measures the difference in the distribution of the detection frames for the existing target detection framework YOLOv5, which ensures that the loss of the small targets is not neglected in the calculation of the forward propagation loss of the model, and we purposefully modify the structure of the feature extraction network in the original framework by adding a feature extraction layer for the detection of the small targets, which enhances the ability of the model to capture the subtle features thus improving the accuracy of small target detection. Through comparative experiments, we evaluated the detection performance of the original YOLOv5 model, the YOLOv5 model using only the NWD loss, the YOLOv5 model with only the additional small-target detection layer, and the full NWD-YOLOv5 model for small-targeted human bodies on the screened COCO2017 public dataset, respectively. The experimental results show that compared to the original YOLOv5 model, the model with only the NWD loss improves the [email protected] metrics on small-target detection by 16.5%, while the model with only the additional small-target detection layer improves it by 19.9%. Finally, the complete NWD-YOLOv5 model achieves the optimal small target detection effect, and its [email protected] index improves by 23.4% compared to the original model, which proves the effectiveness of the proposed method in this paper.