VesselSARNet: Lightweight Ship Detection For SAR Images
Zhenyu Li, Wei Li, Wei Tang, Zhenlong Xu · 2024
Automatic ship detection from Synthetic Aperture Radar (SAR) images represents a pivotal yet formidable challenge. Traditionally, ship detection algorithms have predominantly concentrated on enhancing accuracy, often at the expense of increased computational complexity, rendering them impractical for environments with limited computational resources, such as space-borne processing platforms. To address these limitations, we introduce VesselSARNet, a lightweight and efficient network tailored specifically for computation-constrained scenarios, built upon the deeply supervised object detection (DSOD) framework. VesselSARNet innovatively incorporates a one-shot aggregation strategy and two-way convolution into its foundational backbone, thereby eliminating the redundancy inherent in dense connections of conventional DenseNet architectures by consolidating all features into the final feature maps through a single aggregation step. This approach not only maintains the advantages of concatenation but also significantly enhances the multiplication-accumulate operations (MAC) efficiency and parallel computation capabilities. Furthermore, to bolster the detection capabilities for multi-scale ship targets, VesselSARNet incorporates a multi-scale deconvolution fusion module, which generates highly representative features tailored for small ships. Additionally, we leverage Focal Loss to mitigate the class imbalance issue between foreground (ships) and background, further refining detection performance. Extensive experimental evaluations underscore the effectiveness and superiority of VesselSARNet over state-of-the-art lightweight object detectors, validating its potential for practical applications in resource-constrained environments.