毕业生知识库

       高分辨率遥感图像的海陆分割和舰船检测是遥感图像处理领域的两个研究热点，在海岸线提取、海域交通管制、军事监测等方面有着重要的应用价值。近年来，针对海陆分割和舰船检测的研究工作取得了一定进展。然而，随着遥感图像分辨率的提高，精确的海陆分割和鲁棒的舰船检测越来越具有挑战性，其主要表现在：（1）复杂的陆地特征，导致传统的海陆分割方法在陆地区域存在大量的漏检；（2）海浪、云层和阴影等诸多干扰因素大大增加了海陆分割与舰船检测的难度；（3）如何保持对海陆边界处精细结构的较高分割精度；（4）如何提取具有高鉴别性的舰船特征。

      本文针对传统的海陆分割与舰船检测方法中存在的问题，将经典的图像分割和深度学习方法与高分辨率海陆遥感图像自身特点相结合，提出了几种有效的解决方案。本文的研究内容和主要贡献如下：

1、提出了一种基于自动种子点选择和边界引导图割的海陆分割方法。针对图割模型需要种子点标注的问题，分别提出了监督的种子点学习方法和无监督的多层区域合并方法。在种子点学习方法中，首先通过超像素计算和多特征编码提取训练和测试样本。其中，超像素计算可以充分利用图像的局部信息和减少样本冗余。而融合了光谱、纹理和空间结构信息的多特征可以有效地描述海陆特征差异。然后通过分类器训练对图像中的种子点自动标注。分类器的概率输出使得可以选择更加可靠的种子点。在多层区域合并方法中，利用底层的超像素聚类和基于图模型的快速区域合并方法，并结合海陆亮度差异选择出最大面积的海洋区域。针对图割方法中的“欠分割”问题，引入边界检测信息来降低分割边界通过较长路径的代价，提高了对海陆边界精细结构的分割精度。实验结果验证了种子点选择和海陆分割方法的有效性。

2、提出了一种基于深度卷积神经网络和结构化边界优化的海陆分割方法。在传统的卷积网络结构上做了两点改进。首先，为了抑制海浪等因素的干扰，提高分割结果的平滑性，提出了基于拉普拉斯平滑约束的损失函数。其中对拉普拉斯矩阵的构建采用相邻像素的颜色距离。其次，提出了一种结构化边界优化的方法来进一步增强海陆区域的类内和边界约束。该方法在网络的分割概率图上计算得到海陆边界的概率输出，利用边界的类别和空间结构信息对分割结果进行优化，明显提升了边界分割的准确率。对比实验表明提出的方法优于传统的海陆分割方法和基于深度卷积神经网络的方法。

3、提出了一种基于深度卷积神经网络融合的海陆分割和舰船检测方法。该方法将海陆分割和舰船检测任务置于统一框架内，利用语义分割网络解决港口内的舰船检测问题。首先，在原有的分割网络基础上搭建边界网络形成多任务学习。其中，边界网络通过提取分割网络的层级语义特征来实现。多任务学习为网络的训练提供了多重约束。其次，为了实现分割网络与边界网络的融合，提出了一种基于语义边界的平滑约束，并在反传过程中同时对分割和边界网络进行优化。相比于传统的利用底层颜色信息的约束方法，该方法能更好地利用语义信息提高分割结果的空间一致性，对相邻码头和舰船的分割更加精确。在海陆分割和舰船检测两个数据集上的实验结果验证了所提方法的有效性。

   Sea--land segmentation and ship detection are two prevalent research domains for high-resolution remote sensing images, and have many applications in coastline extraction, sea traffic management and military monitoring, etc.  During last several years, the research on sea--land segmentation and ship detection has experienced great advance. However, with the improvement of the spatial resolution of remote sensing images, there are still some factors make those two tasks challenging. First, due to the complicated intensity and texture distribution of land region, traditional methods often fail to obtain smooth segmentation results. Second, the disturbance of cloud, wave and shadow further adds difficulty for accurate sea--land segmentation and ship detection. Third, the segmentation of some thin and elongated structures around sea-land boundaries is challenging. Finally, more discriminative ship features are needed.

    According to the existing problems in traditional sea--land segmentation and ship detection methods, this paper applies the prevalent segmentation and deep learning methods to high-resolution remote sensing images, and proposes some solutions. The main contributions are as follows:

1. A sea--land segmentation method based on automatic seeds selection and edge directed graph cut (GC) is proposed. The problem of user-specific strokes needed in GC is addressed by providing two automatic seeds selection schemes, namely the seeds learning method and hierarchical region merging method. For seeds learning method, the superpixel is used to replace pixel to extract training and testing samples, which can better employ local information and reduce information redundancy. Then a multi-feature descriptor fuses spectral, texture and spatial information is proposed to encode each sample. The probabilistic support vector machine (SVM) is employed to train the model, and testing samples whose probability outputs are higher than a threshold are selected as seeds. For hierarchical region merging method, the superpixel clustering and graph-based merging method are combined to extract the maximum area of sea region (MASR), which is robust to obtain smoothing results. To reduce the under-segmentation phenomenon in GC, edge prior is integrated in GC model to improve results of some thin and elongated structures around sea--land boundaries. Experiments on sea--land images demonstrate the effectiveness of proposed seeds selection and segmentation methods.

2. A structured edge network for sea--land segmentation is proposed. The convolutional neural network is applied to sea--land segmentation and two innovations are made on top of the traditional structure. First, a Laplacian smooth regularization is proposed to achieve better spatially consistent result. The Laplacian matrix is built with spectral distance of neighboring pixels. Second, a structured edge optimization method is proposed to improve the inner-class regularization of the segmentation network and boundary accuracy of the results by employing both label and structure information of edges. Experiments validate the effectiveness of proposed method by comparing with both traditional sea--land segmentation methods and convolutional neural networks.

3. An edge aware deep convolutional network for sea--land segmentation and ship detection is proposed, which unifies those two tasks into a single framework and applies the semantic segmentation network to parse the harbor image into three typical objects, e.g., sea, land and ship. First, a multi-task model is designed by simultaneously training the segmentation network and edge detection network. Hierarchical semantic features from the segmentation network are extracted to learn the edge network. Second, the outputs of edge pipeline are further employed to refine the entire model by adding an edge aware regularization, which helps to yield very desirable results that are spatially consistent and well boundary located. It also benefits the segmentation of docked ships that are quite challenging for many previous approaches. The proposed method achieves better performance on both harbor image dataset and sea--land image dataset compared with traditional methods.