多尺度特征融合与特征通道关系校准的SAR图像船舶检测

目前深度学习技术在SAR图像的船舶检测中已取得显著的成果，但针对SAR船舶图像中复杂多变的背景环境，如何准确高效地提取目标特征，提升检测精度与检测速度仍存在着巨大的挑战。针对上述问题，该文提出了一种多尺度特征融合与特征通道关系校准的 SAR 图像船舶检测算法。在Faster R-CNN的基础上，首先通过引入通道注意力机制对特征提取网络进行特征间通道关系校准，提高网络对复杂场景下船舶目标特征提取的表达能力；其次，不同于原始的基于单一尺度特征生成候选区域的方法，该文基于神经架构搜索算法引入改进的特征金字塔结构，高效地将多尺度特征进行充分融合，改善了船舶目标中对小目标、近岸密集目标的漏检问题。最后，在SSDD数据集上进行对比验证。实验结果表明，相较原始的Faster R-CNN，检测精度从85.4%提高到89.4%，检测速率也从2.8 FPS提高到10.7 FPS。该方法能够有效实现高速与高精度的SAR图像船舶检测，具有一定的现实意义。 

Ship Detection in SAR Images Based on Multiscale Feature Fusion and Channel Relation Calibration of Features

Deep-learning technology has enabled remarkable results for ship detection in SAR images. However, in view of the complex and changeable backgrounds of SAR ship images, how to accurately and efficiently extract target features and improve detection accuracy and speed is still a huge challenge. To solve this problem, a ship detection algorithm based on multiscale feature fusion and channel relation calibration of features is proposed in this paper. First, based on Faster R-CNN, a channel attention mechanism is introduced to calibrate the channel relationship between features in the feature extraction network, so as to improve the network’s expression ability for extraction of ship features in different scenes. Second, unlike the original method of generating candidate regions based on single-scale features, this paper introduces an improved feature pyramid structure based on a neural architecture search algorithm, which helps improve the performance of the network. The multiscale features are effectively fused to settle the problem of missing detections of small targets and adjacent inshore targets. Experimental results on the SSDD dataset show that, compared with the original Faster R-CNN, the proposed algorithm improves detection accuracy from 85.4% to 89.4% and the detection rate from 2.8 FPS to 10.7 FPS. Thus, this method effectively achieves high-speed and high-accuracy SAR ship detection, which has practical benefits.