Aerial Ship Detection Based on YOLOv5 and Semi-Supervised Learning

Aerial ship detection plays a crucial role in marine resource management and other related fields. Traditional object detection methods rely on a large amount of annotated data, which is costly to obtain. Therefore, this paper proposed a semi-supervised learning-based aerial ship detection method that adopted the Noisy Student strategy. The teacher model predicted unlabeled data, and an adaptive thresholding mechanism based on MC Dropout was employed to refine the confidence of pseudo-labels, ensuring high-quality supervision for training the student model. In the YOLOv5 Backbone, coordinate attention (CA) was introduced by replacing C3 with C3CA to enhance the model's spatial information capture capability. Additionally, in the Neck network, the Bottleneck Blocks in the final C3 module were replaced with Transformer Encoder Blocks (TEB) to improve multi-scale feature fusion. The results showed that the improved YOLOv5 achieved a precision (P) of 90%, recall (R) of 75.7%, and mAP@.5 of 85%. The best-performing semi-supervised model, trained with 30% of the labeled data, achieved a P of 87.3%, R of 69.9%, and mAP@.5 of 80.8%. Although its performance was slightly lower than the fully supervised model, it reduced the annotation workload by 70%, providing a reliable and cost-effective solution for ship monitoring, marine resource protection, and intelligent maritime management.