Efficient ship detection in sar images with dynamic feature smoothing and visual module using omni-dimensional dynamic large-scale convolution

In SAR (Synthetic Aperture Radar) image ship detection tasks, the accuracy of detection is hindered by the presence of a significant amount of clutter and false targets. Ships in these images exhibit diverse scales, which poses a challenge for traditional models to capture the necessary features across these varying scales. Furthermore, the limited texture information available in SAR images complicates the task of distinguishing between targets and the background, leading to suboptimal performance when using conventional deep learning models. To address these challenges, this paper introduces several innovative components. Firstly, a novel dynamic convolution operator is proposed, which allows for adaptively adjusting the smoothing method to mitigate the impact of noise in the data. Additionally, the paper introduces a plug-and-play feature smoothing module as well as a visual module. The feature smoothing module contributes to reducing noise interference, enhancing the overall robustness of the model. Meanwhile, the visual module focuses on augmenting long-range dependency relationships within the data, facilitating a better understanding of the complex interplay between ships and the background. This effectively addresses the issue of texture scarcity in SAR images. Our method was compared with the most advanced Visual module Explicit Visual Center (EVC) using the YOLO7-obb object detection baseline. The experimental results unequivocally demonstrate that our proposed method achieved a notable 1.2% improvement in the Average Precision (AP) value.