BD-YOLO: detection algorithm for high-resolution remote sensing images

Abstract

This paper focuses on detecting extremely small targets in aerial images. Compared to common datasets, the average size of targets in remote sensing images is only 12.8 pixels, significantly smaller than those in common datasets. Therefore, directly applying existing detectors to aerial images is ineffective. To address this issue and ensure real-time performance, This paper propose BD-YOLO, which incorporates five key innovations. A dual Backbone route was proposed to maintain data integrity and achieve high-resolution aerial remote sensing detection. Additionally, a new feature fusion method was developed to comprehensively merge shallow and deep information. To supplement small-sized target information, a new network structure was proposed. The detector strategy used by BD-YOLO considers the detection accuracy of objects with different sizes. Furthermore, a lightweight method was adopted to ensure real-time performance of the algorithm. BD-YOLO outperformed YOLOv8s on the AI-TOD dataset, achieving a higher mAP by 2.4%. Similarly, on the Visdrone dataset, BD-YOLO achieved a 2.5% higher mAP compared to YOLOv8s. Additionally, on the Tinyperson dataset, BD-YOLO achieved a 0.6% higher mAP than YOLOv8s. Notably, BD-YOLO maintains real-time performance while ensuring accurate object detection.