An Improved Object Detection Algorithm for UAV Images Based on Orthogonal Channel Attention Mechanism and Triple Feature Encoder

Abstract

Object detection in Unmanned Aerial Vehicle (UAV) imagery plays an important role in many fields. However, UAV images usually exhibit characteristics different from those of natural images, such as complex scenes, dense small targets, and significant variations in target scales, which pose considerable challenges for object detection tasks. To address these issues, this paper presents a novel object detection algorithm for UAV images based on YOLOv8 (referred to as OATF-YOLO). First, an orthogonal channel attention mechanism is added to the backbone network to imporve the algorithm's ability to extract features and clear up any confusion between features in the foreground and background. Second, a triple feature encoder and a scale sequence feature fusion module are integrated into the neck network to bolster the algorithm's multi-scale feature fusion capability, thereby mitigating the impact of substantial differences in target scales. Finally, an inner factor is introduced into the loss function to further upgrade the robustness and detection accuracy of the algorithm. Experimental results on the VisDrone2019-DET dataset indicate that the proposed algorithm significantly outperforms the baseline model. On the validation set, the OATF-YOLO algorithm achieves a precision of 59.1%, a recall of 40.5%, an mAP50 of 42.5%, and an mAP50:95 of 25.8%. These values represent improvements of 3.8%, 3.0%, 4.1%, and 3.3%, respectively. Similarly, on the test set, the OATF-YOLO algorithm achieves a precision of 52.3%, a recall of 34.7%, an mAP50 of 33.4%, and an mAP50:95 of 19.1%, reflecting enhancements of 4.0%, 3.3%, 4.0%, and 2.6%, respectively. To further validate the model's robustness and scalability, experiments are conducted on the NWPU-VHR10 dataset, and OATF-YOLO also achieves excellent performance. Furthermore, compared to several classical object detection algorithms, OATF-YOLO demonstrates superior detection performance on both datasets and indicates that it is better suited for UAV image object detection scenarios.