DE-YOLO: Detail-Enhanced Maritime Object Detection Algorithm Based on YOLOv8

When UAVs are applied to maritime object detection, the high occurrence frequency of small objects, object occlusion and complex environments can lead to a reduction in recognition accuracy. To address this problem, we propose a DE-YOLO algorithm based on detail-enhanced convolution. The algorithm first adds an extra detection head to YOLOv8s for recognizing very small objects. Then DEConv is integrated into the original backbone modules Conv and C2f, enabling the backbone modules to utilize both gradient and intensity information to generate more discriminative feature maps. In addition, DE-YOLO uses PIoUv2 to solve the problem of slow convergence due to the phenomenon of region enlargement, accelerating the regression process. Experimental results on the open water object detection dataset show that our algorithm achieves 45.3% mAP, 79.5% AP50 and 39.4% APsmall detection accuracy, which are 7.3%, 15.8% and 14.0% higher than the baseline, respectively.