YOLOv8-IPAN: Enhancing Ship Detection in Maritime Surveillance with Integrated Pyramid Attention and Progressive Feature Fusion

In today's rapidly evolving shipping industry, ship identification and detection technologies have increasingly become core development areas for key sectors such as maritime transportation management and ship traffic supervision. However, practical applications face several challenges when ships are used as detection objects. These include issues such as the non-uniform pixel distribution of target ships in images, the high similarity in shapes of ships across different categories, and mutual occlusion between targets. From a technical perspective, there are two core challenges. First, when dealing with small or incomplete targets, the downsampling process of the feature map results in the loss of high-level semantic features and feature degradation, which leads to a significant decline in detection performance. Second, existing feature extraction techniques often struggle to accurately capture and differentiate subtle variations in the characteristics of different ship categories, particularly in scenarios where targets are densely packed and exhibit similar shapes. This issue is particularly evident in such crowded settings. To address these challenges, this paper introduces an innovative architecture, the YOLOv8-Integrated Pyramid Attention Network (YOLOv8-IPAN), based on the YOLOv8n optimization. The proposed architecture first integrates an Asymptotic Feature Pyramid Network (AFPN) to enhance the fusion of multi-level features [1]. Additionally, we incorporate a lightweight Convolutional Block Attention Module (CBAM) into the YOLOv8n framework. To validate the performance of the YOLOv8-IPAN architecture, comprehensive testing, and training were conducted using the SeaShips public maritime dataset, achieving impressive results. The experimental data show that the YOLOv8-IPAN model achieves 97.4% accuracy, 97.6% regression rate, and 99.0% mAP50, demonstrating its effectiveness and strong generalization ability. These results provide strong evidence supporting the feasibility of the proposed method.