MSCK-Net: Multiscale Chinese Knot Convolutional Network for Dim and Small Infrared Ship Detection

초록

Remote sensing infrared ship detection is crucial for maritime safety and traffic management in civil-military integrated applications. However, the detection accuracy and robustness of current methods are still limited by dataset constraints, including small scale, narrow target size distribution, sparse targets, and high scenario specificity. To address these problems, we integrate publicly available datasets to construct IRShip—a relatively large-scale infrared ship detection dataset comprising 27,138 images, significantly improving data scale and diversity. We further design a copy-poisson blend (CP-PB) offline data augmentation approach and introduce a dense one-to-one (Dense-O2O) online augmentation strategy, which improve adaptability to complex backgrounds and scale variations, mitigate problems from sparse targets, and improve dataset utility and model robustness. We propose MSCK-Net, a multiscale chinese knot convolutional network tailored for detecting dim and small infrared ship targets (DSIRST). Specifically, we propose a novel Chinese Knot Convolution (CKConv) that better aligns with both the local features of small targets and the morphological characteristics of ship targets, thereby significantly enhancing low-level feature representation. Built with CKConv, the multiscale knot block (MSK-Block) and Stem-ck modules enhance deep feature transmission efficiency and global modeling of DSIRST, leading to notable gains in detection performance. Extensive experiments on IRShip, NUDT-SIRST-Sea, ISDD, IRSDSS, and Maritime-sirst demonstrate that MSCK-Net-M achieves state-of-the-art performance, with AP<inf>50</inf>, AP<inf>75</inf>, and AP of 82.5%, 53.1%, and 50.9% on IRShip, significantly outperforming the existing 20 general object detectors and 6 infrared target detectors. Generalization experiments on four sub-datasets further verify the effectiveness and robustness of our proposed method. The code and dataset are available at: https://github.com/sjxxrh/MSCK-Net. © 1980-2012 IEEE.

키워드