An Improved YOLOv3 Airplane Target Detection Algorithm Based on Multi-Scale Feature Fusion

The detection and identification of airplane targets images have essential application value in remote sensing. Being aimed at the problem of unbalanced accuracy in the current detection of small-scale airplane targets, an improved airplane target detection algorithm based on YOLOv3 is proposed. First, with the feature pyramid network of the YOLOv3 model redesigned, a prediction feature layer for small targets was added, and the features of the external network and the deep network were fused to improve the model’s detection ability small targets. Then, the K-means clustering algorithm is utilized to adjust the a priori box scale of the prediction feature layer. Lastly, experimental works based on the airplane target detection show: that the average precision of the YOLOv3 algorithm of the proposed method achieves 95.3%, which is 22.9% higher than the original algorithm. The precision and recall rates are improved, and the PFS is 30.86. These results prove the effectiveness of the method proposed in this letter.