基于深度协同稀疏编码网络的海洋浮筏SAR图像目标识别

浮筏养殖广泛存在于我国近海海域, 可见光遥感图像无法完全准确地获取养殖目标, 而基于主动成像的合成孔径雷达(Synthetic aperture radar, SAR)遥感图像能够得到养殖目标, 因此采用SAR图像进行海洋浮筏养殖目标识别. 然而, 海洋遥感SAR图像包含大量相干斑噪声, 并且SAR图像特征单一, 使得目标识别难度较大. 为解决这些问题, 提出一种深度协同稀疏编码网络(Deep collaborative sparse coding network, DCSCN)进行海洋浮筏识别. 本文方法对预处理后的图像先提取纹理特征和轮廓特征, 再进行超像素分割并将同一个超像素块特征组输入该网络进行协同表示, 最后得到有效特征并分类识别. 通过人工SAR图像和北戴河海域浮筏养殖SAR图像的实验验证所提模型的有效性. 该网络不仅具有优异的特征表示能力, 能够获得更适合分类器的特征, 而且通过近邻协同约束, 有效抑制相干斑噪声影响, 所以提高了SAR图像目标识别精度.

Research on Marine Floating Raft Aquaculture SAR Image Target Recognition Based on Deep Collaborative Sparse Coding Network

Floating raft aquaculture is widely distributed in the offshore ocean of China. Since raft information cannot be obtained accurately in the visible remote sensing image, active imaging images acquired from synthetic aperture radar (SAR) are applied. However, oceanic SAR images are seriously contaminated by speckle noise, and effective features of SAR images are deficient, which make recognition difficult. In order to overcome these problems, a deep collaborative sparse coding network (DCSCN) is proposed to extract features and conduct recognition automatically. The proposed method extracts texture features and contour features from the pre-processed image firstly. Then, it segments the image into patches and learns features of each patch collaboratively through the DCSCN network. The optimized features are used for recognition finally. Experiments on the artificial SAR image and the images of Beidaihe demonstrate that the proposed DCSCN network can accurately obtain the area of floating raft aquaculture. Since the network can learn discriminative features and integrate the correlated neighbor pixels, the DCSCN network improves the recognition accuracy and has better performance in overcoming the contamination of speckle noise.