GDP-YOLO:A Lightweight Nighttime Infrared Ship Detection Network

Infrared ship target detection plays a crucial role in maritime applications. However, existing methods often struggle with the trade-off between detection precision and computational efficiency, due to small target sizes in low-contrast images and the resource constraints of embedded platforms. To address this issue, we propose a lightweight yet high-precision network, the GDP-YOLO, based on YOLOv11n. Firstly, the group shuffle convolution module is presented for feature extraction; therefore, the computational complexity of the respective modules is reduced while discriminative features are preserved through channel grouping and feature rearrangement techniques. Secondly, a dynamic feature fusion Module is presented in the feature fusion step, by utilizing a global context-guided mechanism to adaptively fuse multiscale image features. The comprehensive representation ability of our networks is enhanced in both local features and semantics. Finally, a P2 head is utilized in the detection; The redundant P5 detection head is removed, therefore, the algorithm's detection ability for low contrast and small-sized ship targets is improved. The comparative experimental results demonstrate that the proposed GDP-YOLO outperforms the comparative methods, particularly achieving improvements of 2.8%, 0.9%, and 3.9% in mAP@0.5, mAP@0.5\((\sim)\)0.95, and Recall, respectively, compared to YOLOv11n. Furthermore, our GDP-YOLO decreased to a parameter size of only 4.1 MB with an inferring speed of 200.45 frames per second. Our source code will be publicly available at \href{https://github.com/hnczwj2008/GDP-YOLO/}{GDP-YOLO}.