基于邻域重构的DEM非局部超分辨率方法研究

Digital terrain model plays an important role in many applications. To obtain high accuracy terrain model, two ways are now used. The first is to capture more accurate and dense DEM using modern equipement. Another is to introduce new approachs to decrease the terrain reconstruction model for a certain DEM. The former is direct and costly. The defect of the second one lies in it can't greatly improve the terrain accuracy. Considering this fact, we propose a new method which is to do super resolution (accuracy) on basis of the original low accuracy DEM and part of high accurate data. The accruarcy afer super resolution should be among the original data and the newly ontained DEM. Moreover, we try to make the accuracy be high. Traditionally, the super resolution methods consider the non local similarities on the whole regions, establish the relationship between the region to be processed and the DEM in training data set, then mapping the high resolution information to the low resolution data. Compared to images, the terrain varies gently. Then, a sub-region should be highly realted to its neighbors. Inspired by those facts, we propose a new super resolution method which is based on the supported relationship among neighbors and the non local similarities. This new strategy considers the local relationship in neighbors and the similarities among non local sub-regions at the same time. It mainly includes five steps. The first is to find the similar sub-regions utilizing translation, scale and rotation invariant character description algorithms. Based on the characters, some similar sub-regions can be found. The second is to research the local reconstruction method for sub regions which tries to decrease the reconstruction error. On basis of non local similarity and the local relationship, the third part is to mapping the high resolution information to the low resolution data. Then, how to estimate the accuracy of result will be discussed, in which, the relationship between the training data and the accuracy after super resolution is the crucial problem. The last one is to accelerate the computation using the GPU. According to the related theory and current experiments, the strategy is advanced and feasible.

高精度地形模型在国民经济中具有重要的应用价值，DEM是建立地形模型的基础数据。通过外业测量获取高精度的DEM；基于DEM的不确定性，建立相应的曲面重建方法，减少或控制DEM误差引起的重建误差，是目前提高地形曲面精度的主要手段。前者生产成本高，后者对精度的提高相对有限。为此，本项目提出一种新的策略，在原始数据基础上，通过测量手段获取目标区域部分高分辨率/精度的DEM，再对两组数据进行联合处理，生成精度介于二者之间的DEM，并尽可能接近精度上限。该策略意图在一定的外业工作量基础上，通过发掘不同分辨率数据的冗余性和互补性，以提高DEM的精度，研究意义十分明显。目前，未见有与之相同的研究报道，本项目在借鉴图像超分辨率方法的基础上，技术思路是：基于子区域相似性和邻域相容性，在原始数据中，寻找待超分辨率数据的非局部相似子区域，根据相似程度，将高分辨率数据映射到对应的区域，从而实现DEM的超分辨率。

DEM广泛应用于人们的日常生活和生产中，人们对高分辨率/精度DEM的追求是无止境的。传统提高DEM分辨率的方法是利用高精度的仪器进行多次测量，而本项目提出了DEM超分辨率问题和方法，策略是基于一定的高分辨率DEM样本，利用非局部、邻域重构和深度学习等方法构建高、低分辨率的映射关系，以降低获取高分辨率DEM的成本。.主要研究内容一，基于非局部的DEM超分辨率，设待处理的DEM具有部分高分辨率的数据，与之对应的低分辨率DEM组成学习样本，非局部策略则将根据不同分辨率DEM非局部关系的一致性，重构高分辨率的DEM。内容二，研究了逐级超分辨率的思路，并非直接将DEM/图像直接提高到目标分辨率，而是逐级进行，这种策略找到的相似样本稳定可靠，能提高结果的鲁棒性。内容三，研究了基于深度学习的DEM超分辨率，采用CNN构建高、低分辨率DEM的映射关系，由于大量高分辨率的DEM样本获取比较难，而图像样本则容易获得，同时考虑到DEM与图像动态范围差异大，于是，提出了利用图像梯度作为学习样本训练CNN；应用时，CNN的输入是DEM的梯度，输出是其对应的高分辨率的DEM梯度图，即此方法对DEM的梯度而非DEM本身进行超分辨率，再利用超分辨率的梯度和原始DEM，基于局部邻域重构关系重建高分辨率的DEM。内容四，基于梯度增强的超分辨率的研究，其基本的分类器是回归树，设计了一种错误修正算法迭代地进行树的训练，即回归树是通过梯度提升的方法进行串行训练，所有增强树合并起来形成精确分类器，建立高低分辨率DEM/图像的映射关系。.DEM超分辨率是提高DEM精度的有效方法，与基于非局部策略相比，基于CNN超分辨率方法的鲁棒性强、精度高。根据绝对值差、坡度和坡向等DEM误差统计指标，基于CNN的梯度超分辨率方法，精度明显优于双三次和Kriging等方法。.本项目提出了一种通过数据处理提高DEM分辨率/精度的方法，利用图像样本的梯度进行CNN训练，再重建高分辨率DEM，有效地模拟了人们对不同分辨率场景关系的重构过程，克服了高分辨率大量DEM样本获取的困难，为分辨率的提高提供了新的思路。