YOLO-FCNET: Enhancing SAR Ship Detection with Fourier Convolution in YOLOv8

Zihao Zhang,Ying Li
DOI: https://doi.org/10.1007/978-981-97-6125-8_15
2024-01-01
Abstract:Detecting ships in Synthetic Aperture Radar (SAR) images is crucial for maritime traffic surveillance and safety.Despite the extensive application of deep learning in SAR ship detection, accurately identifying ships, especially smaller ones, remains a significant challenge. The unique features of SAR images, including noise interference and resolution constraints, add complexity to the detection task. In this paper, we introduce YOLO-FCNET, a novel ship detection approach that leverages the frequency domain analysis based on YOLOv8, to enhance both the precision and efficiency of SAR ship detection.We begin by incorporating the FCSE Block, a module that enriches the model's feature representation by intelligently amplifying salient features through a channel squeeze-excitation mechanism, while muting less informative ones, thus sharpening the model's ability to identify critical ship attributes.Next, we introduce the Fourier Unit module, which empowers our model to discern periodic patterns and textures within the frequency domain via Fourier Convolution. This capability aids in filtering out noise and background clutter, thereby enhancing the extraction of meaningful signals.Finally, the Spectral Transform module is integrated to achieve a harmonious fusion of multi-scale and multi-feature representations. This module employs Spectral and Local Fourier Convolution to bolster the model's capacity to detect ships from various angles and sizes, and to adapt to diverse scale and shape variations.Experiments on a well-established SAR ship detection dataset (SSDD) have confirmed the efficacy of our proposed method. The results demonstrate that YOLO-FCNET significantly outperforms existing methods, particularly in the detection of smaller ships.
What problem does this paper attempt to address?