Sentinel-2卫星的多光谱轻量级船舶目标检测算法

近年来深度卷积神经网络在可见光船舶检测方面取得了显著的进展，然而，大多数相关研究是通过改进大型的网络结构来提高检测性能，因此加大了对更高计算机性能的需求。此外，可见光图像难以在云、雾、海杂波、黑夜等复杂场景检测到船舶。针对以上问题，提出了一种融合红（red, R）、绿（green, G）、蓝（blue, B）和近红外（NIR）4个波段光谱信息的由粗到精细的轻量型船舶检测算法。与现有的方法中根据光谱特性利用水体检测算法提取水体区域不同之处是该算法是利用改进的水体检测算法来提取船舶候选区域。为获取更准确的候选区域，对船舶、厚云、薄云、平静海面、杂波海面5种场景中4个波段的像素值进行了统计分析，选取近红外大于阈值作为辅助判断，并以其中心点获取候选区域32×32大小的切片，并对切片进行非极大值抑制，由此获得了船舶粗检测结果。随后构建了轻量级LSGFNet网络对船舶候选区域切片进行精细识别。构建的网络融合了1×1卷积提取的波谱特征与3×3的提取几何特征，为防止光谱特征与几何特征的信息在融合时“信息不流通”，在LSGFNet网络中引入了ShuffleNet中的通道打乱机制，并减小了模型结构，与典型的轻量级网络相比具有更好的效果且模型较小。最后，利用Sentinel-2卫星多光谱10 m分辨率数据构建了512×512大小的1 120组数据进行粗检测，以及32×32大小的6 014组数据进行精细网络训练，其中候选区域粗提取的查全率为98.99%，精细识别网络精确度为96.04%，不同场景下的平均精确度为92.98%。实验表明该算法在抑制云层、海浪杂波等干扰的复杂背景下具有较高的检测效率，且训练时间短、计算机性能需求低。

Multispectral Lightweight Ship Target Detection Algorithm for Sentinel-2 Satellite

Recently, deep convolutional neural networks have made remarkable progress in visible light ship detection. However, most improve detection performance by improving large network structures, which require strong computer performance. In addition, the visible image is difficult to detect ships in the cloud, fog, sea clutter, night and other complex scenes. In order to solve these problems, this paper proposes a lightweight ship detection algorithm, which integrates the spectral information of Red, Green, Blue and NIR bands from coarse to fine. This paper uses the improved water detection algorithm to extract the ship candidate area from the existing methods that extract the water area by using the water detection algorithm according to the spectral characteristics. To obtain a more accurate candidate area, in this paper, the ships, the thick clouds, cloud, calm sea, five kinds of cluttered sea four bands of pixels in the scene have carried on the statistical analysis. Near-infrared is greater than the threshold as the auxiliary judgment, and in its center for candidate area the size of 32×32 slices, and to the maximum inhibition of slice, thus obtained the coarse detection results of the ship. Then constructs a lightweight LSGFNet network for fine identification of ship candidate regions. In the structural design of the network, the spectral features extracted by 1×1 convolution and the geometric features extracted by 3×3 are fused. In order to prevent the “information not flowing” during the fusion of spectral features and geometric features, the channel disruption mechanism in ShuffleNet is introduced in the LSGFNet network, and the model structure is reduced. Compared with the typical lightweight network, it has a better effect and a smaller model. Finally, 1 120 sets of data with 512×512 and 6014 sets of data with 32×32 were constructed for rough detection and fine network training using sentinel-2 multi-spectral 10-meter resolution data. Among them, the recall rate of rough extraction of candidate regions was 98.99%, and the fine identification network precision was 96.04%. The overall average precision is 92.98%. Experimental results show that the proposed algorithm has high detection efficiency in the complex background of suppressing clouds, sea clutter and other disturbances, the training time is short, and the computer performance demand is low.