In a study published in Pattern Recognition on September 29, researchers from the Changchun Institute of Optics, Fine Mechanics and Physics of the Chinese Academy of Sciences developed a new feature extraction network for ship detection in optical remote sensing images. The method, called Mamba-CWF, combined global scene understanding with detailed shape recognition to better identify ships in complex backgrounds. It achieved mean average precision (mAP) scores of 90.73%, 70.4%, and 78.2% on three widely used remote sensing datasets, while maintaining relatively low computational complexity.

Ship detection in satellite and aerial imagery is important for maritime monitoring and vessel traffic management. However, ships can appear very small, long and narrow, or at different orientations, and are often surrounded by ports, coastlines, and other visually similar structures. Existing models therefore need to understand both the overall scene and fine details such as object boundaries and shapes. 

Transformer-based approaches can capture broad contextual information but require substantial computation, while Mamba offers a more efficient alternative. Yet directly converting a two-dimensional image into a one-dimensional sequence can disrupt the spatial relationships needed to recognize ship structures accurately.

To address this problem, the researchers developed Mamba-CWF with three complementary components. A Global Visual Clustering Scan reorganized visually related image features before sequence modeling, helping the model retain object boundaries while understanding the wider scene. A lightweight Local Shape-Adaptive Feature Extraction branch then strengthened the representation of elongated structures and directional details typical of ships. A Wavelet Attention Enhanced Fusion module combined global and local information in both spatial and frequency domains. In simple terms, the network was designed to see both the "big picture" and the fine contours of a ship without requiring excessively expensive computation.

The researchers evaluated Mamba-CWF on the HRSC2016, DIOR-R, and DOTA benchmarks. On DOTA, it achieved 78.2% mAP and outperformed commonly used ResNet50, Swin-T, and VMamba-T backbones. On DIOR-R, it achieved 70.4% mAP and the best results in 11 object categories, showing particular advantages for elongated and orientation-sensitive targets such as ships, bridges, and harbors. Across several detection frameworks, its floating-point operations were only 29.9% of those required by ResNet50.

The study provides an efficient way to adapt Mamba-style sequence modeling to two-dimensional remote sensing images while preserving spatial and shape information. The method also maintained relatively stable detection performance when images were affected by noise, blur, compression, and reduced resolution, indicating potential for practical remote sensing applications where image quality may vary. Such capabilities could support more reliable maritime monitoring from satellites and aerial platforms. 

Further work will focus on reducing model size through lightweight design, pruning, and quantization to improve deployment on resource-constrained devices.