Research on small target detection in aerial images based on Yolov7 and Transformer

With the further development of the network, it has become more rare to obtain effective feature information of small targets in aerial images, and the resulting detection accuracy has become more challenging. To solve this problem, this paper combines the target detection algorithm YOLOv7, and introduces the neighborhood attention transformer (NAT) module based on it. A small target detection method based on Yolov7 and Transformer for aerial images is proposed. Compared with the traditional CNN module, the local feature Transformer at any two locations can interact and correlate, obtain more context information and capture more global information. By introducing location offset information, NAT also retains the advantages of visual Transformer, and further improves the effect of local feature extraction. In addition, considering the limited computing resources of edge computing devices, this paper adopts Light-weight Convolutional Encoder Network (LCEN) and introduces an embedded hierarchical attention network in the feature extraction stage. To balance the detection accuracy and inference speed of the model. The experimental data fully prove that the proposed method is effective in solving the problem of small target detection in aerial images. In addition, when compared with other mainstream aerial small target detection models on the two datasets, the proposed method shows higher average accuracy. It is especially worth mentioning that compared with the latest aerial scene target detection model TPH-YOLOv5, the average accuracy of the proposed method is improved by 7.34%.