LTGS: an optical remote sensing tiny ship detection model

• Published in: Pattern analysis and applications : PAA Vol. 28; no. 3
• Main Authors: Yan, Chunman, Qi, Ningning
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2025; Springer Nature B.V
• Subjects: Accuracy; Computer Science; Datasets; Modules; Object recognition; Original Article; Parameters; Pattern Recognition; Remote sensing; Ships; Target detection; Deep learning; Ship detection; Tiny target detection; Optical remote sensing; YOLOv8
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-025-01503-z

In optical remote sensing (ORS) ship detection tasks based on deep learning, traditional models have made significant progress in detecting small ships, but still face challenges such as insufficient accuracy and missed detection for tiny ships. To address this problem, we propose LTGS (Lightweight Tiny-target Guided Sensing), a lightweight and optimized detection model based on YOLOv8n, specifically designed for tiny-object detection in ORS imagery. The design of LTGS is guided by a novel architectural framework termed the Detection Path Optimizer (DPO), which redefines the detection process as a coordinated multi-path optimization system to improve detection performance and efficiency. LTGS incorporates three key components. First, the Lightweight Cascaded Perception Module (LCPM) replaces the original C2f in the backbone to enhance feature extraction and gradient propagation through hierarchical cascaded convolutions and identity mappings, while significantly reducing parameters. Second, the Global Spatial Pyramid Pooling-Fast (GSPPF) module replaces the traditional SPPF to facilitate global-local semantic fusion via adaptive pooling, improving robustness in cluttered environments. Third, the Tiny Target Detection Head (TD) introduces guided upsampling and multi-path fusion to restore fine-grained details lost during downsampling, thereby boosting localization accuracy for tiny targets. Experiments on the public datasets HRSC2016 and MASATI v2, along with the Self-managed and Height-managed datasets, verified the model’s effectiveness. The results show that the improved model increased mAP@50 by 2.1%, reduced parameters by 0.3 M, and reduced the weight file size by 0.3 MB, with particularly superior performance in tiny ship detection tasks.