Improved algorithm for small object detection in aerial image based on YOLOv7 model

For the problems of high false negative and false positive rate of the current detection technology for small object detection in aerial image, we proposed a small object detection for aerial image based on the improved YOLOv7. Firstly, a SPPCSPC-MP model is proposed to capture weakly salient small objects in aerial images in order to improve the model's feature capture capability for small objects. Secondly, we introduced the SPD-Conv module for low-resolution images and small target refinement detection. The SPD-Conv module solves the problems of losing fine-grained information and low learning efficiency for feature representation in the original model. Then, The K-means++ is used to re-cluster the anchor boxes, which makes the anchor boxes fit the objects more closely and increases the network's localisation accuracy on the object. Finally, we evaluate the performance of the improved algorithm on the processed DOTA aerial image dataset. The experimental results show that the improved YOLOv7 model has a mean average precision (mAP) of 77.8% and a recall of 75.2%, which is an improvement of 5.3% and 5.6% compared to the baseline model, and effectively improve the quality of small object detection in aerial image.