Aircraft Target Detection in Satellite Remote Sensing Images Based on Improved YOLOv5

Abstract

Compared with natural images, remote sensing targets have small and dense target shape and complex background, resulting in low detection accuracy and inaccurate identification of target positions. To better extract the features and locations of aircraft targets, this paper proposes a YOLOv5-absorbed algorithm based on the YOLOv5 algorithm. The YOLOv5-absorbed algorithm removes the low-resolution feature layers of the Backbone and the Neck and prunes the prediction head to reduce the loss of position information. At the same time, a new up-sampling module is added to enlarge the feature map in the PAN (Path Aggregation Network) and improve the detection accuracy of aircraft targets in remote sensing images. On this basis, the Coordinate Attention mechanism is introduced to make the network pay attention to a larger area, and DIOU NMS (Distance IoU Non-Maximum Suppression) is introduced to improve the detection accuracy of dense targets. The experimental results of the test data set show that compared with the YOLOv5 algorithm, the YOLOv5-absorbed algorithm has a faster convergence speed and smaller loss, mAP (mean Average Precision) increased from 89.6% to 95.3% and the number of parameters decreased from 92.216M to 36.046M.