Abstract:

Aiming at the problem of pixel classification error and boundary segmentation blur caused by the complex shape of land and dock, and the complex background in the sea-land semantic segmentation, this paper proposes a new method based on deep convolutional neural network for sea-land segmentation of remote sensing image. This method implements the pixel-by-pixel classification of the target in an end-to-end training manner.  Aiming at the problem of pixel classification error in sea-land segmentation, this paper designs three parallel coding structures with different scale images as input.  It can enrich the semantic information of feature representation and increase the pixel classification accuracy by merging the feature maps of different scales.  Aiming at the problem of boundary segmentation blurring in sea-land segmentation, this paper designs decoding structure which merges more fine position information in low coding layer structure to achieve a more accurate up-sample operation, restore dense location information, and improve the accuracy of sea-land segmentation.  At the same time, in order to effectively verify the network framework proposed in this paper, this paper constructs a new sea-land segmentation dataset HRSC2016-SL.  Compared with the latest semantic segmentation algorithms, the proposed algorithm achieves better segmentation results.

Key words: convolutional neural network, sea-land segmentation, semantic segmentation, remote sensing image

摘要：

针对海陆语义分割中陆地、码头形状多样，背景目标复杂等情况造成的像素分类错误、边界分割模糊等问题，提出了一种新的基于深度卷积神经网络的遥感图像海陆语义分割方法。该方法以端对端的训练方式实现了对目标的逐像素分类，为了解决海陆分割中像素分类错误，设计以不同尺度图像为输入的三个并行的编码结构，通过融合不同尺度的特征图，丰富特征代表算子的语义信息，增大像素分类准确率。为了解决海陆分割中边界分割模糊，通过设计能够融合编码结构中低层精细位置信息的解码结构，对特征图进行更加精确的上采样，恢复像素的密集位置信息，提高海陆分割准确度。为有效验证所提网络框架的优势，构建了海陆分割数据集HRSC2016-SL进行算法性能比较。与最新的语义分割算法相比，所提算法取得了更好的分割结果。

关键词: 卷积神经网络, 海陆分割, 语义分割, 遥感图像