语义增强与高阶强交互的SAR图像舰船检测

Semantic Enhancement and High-order Strong Interaction for Ship Detection in SAR Images

合成孔径雷达(synthetic aperture radar,SAR)图像背景信息复杂、舰船目标边缘模糊,且多为容易丢失的小尺度舰船目标。针对上述问题,提出语义增强与高阶强交互的SAR图像舰船检测。该方法利用部分卷积与非对称卷积构建部分非对称卷积聚合网络,在减少计算复杂度、轻量化主干网络的同时,更好地捕捉多尺度舰船特征,同时在上采样部分引入双层路由注意力,增强对图像上下文信息的利用。另外,通过递归的方式进行特征提取,可以较好解决区域内信息交互的问题,实现不同级别特征之间的高阶交互建模,提升模型检测能力。在公开的HRSID遥感数据集上进行实验的结果表明,该方法的检测精度达到91.23%,相比原模型提升5.13%,准确率与召回率分别提升2.41%和7.16%,与主流算法相比具有较好的检测效果。

The background information in synthetic aperture radar(SAR)images is complex,and the edges of ship targets are often blurred,making it difficult to detect small-scale ship targets that are prone to be missed.To address these issues,this paper proposes a SAR ship detection method that combines semantic enhancement and high-order strong interactions.By using partial convolution and asymmetric convolution,a partially asymmetric convolutional aggregation network is constructed.This network reduces computational complexity and lightweight the backbone network,while better capturing multi-scale ship features.In addition,a dual-path routing attention mechanism is introduced in the upsampling part to enhance the utilization of contextual information in the image.Feature extraction in a recursive manner can better solve the problem of information interaction in the region,realize high-order interactive modeling between different levels of features,and improve model detection capabilities.Experimental results on the publicly available HRSID remote sensing dataset demonstrate that the improved method achieves a detection accuracy of 91.23%,which is a 5.13%improvement over the original model.The precision and recall rates are improved by 2.41%and 7.16%,respectively.Compared to mainstream algorithms,the proposed method achieves better detection performance.