基于不规则区域信息提取的扫描地形图图像分割算法研究

Topographic map segmentation technology is the foundation of geographic information extraction from topographic maps. Moreover, it is an important component in methodology of geographic information data acquisition. There were numerous previous works on this aspect, but most of them were either inefficiency or lacking of accuracy. This project aims at solving the challenges in scanned topographic map segmentation, which based on the model of feature extraction from irregular region. This project includes four proposed topics as following. The first one is semi-automatic labeling and pre-segmentation techniques, which will obtain accurate labeling results with a few labors as well as generating appropriate superpixels. The second one is segmentation based on deformable convolutional neural networks, on both pixel and superpixel levels. The deformable convolutional neural networks will precisely extract the features from irregular geographic elements, to help segmenting scanned topographic maps accurately. The third one is based on information fusion at multiple granularity levels, which can accurately predict each pixel without information loss. The last one is segmentation based on irregular superpixel classification, which employs the random mapping selection to obtain a pixel sequence from each superpixel. With this unifying, superpixels with different shapes and sizes can be classified with the features extracted from neural networks. With the researches of above topics, we will propose a series of segmentation models, which makes the segmentation performance efficient and accurate.

扫描地形图分割技术是地形图地理信息提取的重要基础，也是空间地理信息数据获取方法的重要组成部分。然而目前的分割算法普遍存在分割效率低、分割准确性差等缺点，本项目旨在对扫描地形图分割这一领域进行深入研究，建立基于不规则区域特征提取的图像分割框架。拟开展的研究内容包括：研究扫描地形图半自动地理要素标注及预分割技术，从而大幅度减少人工标注工作量，并得到准确的超像素划分；研究可变形卷积神经网络模型，在像素级以及超像素级分别利用可变形卷及网络对不规则的地理要素进行特征提取，从而实现准确分割；研究跨粒度信息融合的精细分割模型，真正意义上实现对每一个像素进行无信息损失的精确分类；研究基于不规则超像素的扫描地形图分割模型，利用随机抽样像素序列映射等思想，对大小形状不一致的超像素进行深层特征提取和分类。通过以上研究，提出一系列应对扫描地形图中要素分布复杂特点的分割模型，实现扫描地形图高效、准确的分割。

扫描地形图分割技术是地形图地理信息提取的重要基础，可以为研究生态变化、地表覆盖变化以及人口流向和城市化等课题提供数据支撑。然而目前的分割算法普遍存在着分割效率低、分割准确性差等缺点，基于此，本项目旨在对扫描地形图分割这一领域进行深入研究，建立基于深度神经网络模型的图像分割框架。本项目在执行阶段提出了一种基于区域色彩与空间信息的地形图中线要素分割算法，该算法定义了色彩混乱程度的概念，并基于这一概念对混合色区域和纯色区域分别划分，从而实现分割；提出了一种新的基于超像素和浅层卷积神经网络的扫描地形图分割算法SSCNN，该算法利用超像素的思想克服扫描地形图中固有的假彩色、渐变色、混淆色等干扰，利用对超像素的分类代替直接对像素分割；提出了一种用于图像分割的神经网络结构ABCNet，提出了一种可调节的Soft Dice Loss代价函数，可以提取图像不同尺度下的特征信息，并以极少的参数有效地融合不同层次的特征。通过对以上分割算法在自建数据集上的验证，在目前已经报道的相关领域中，均达到了最好的彩色扫描地形图分割结果。这些研究的开展，将会进一步开拓彩色扫描地形图地理信息提取技术的应用，研究成果可用于军事、民用等相关领域，具有较好的理论和应用价。