大数据科学下基于条件随机场和神经网络模型的多源图像融合

As a critical part of big data analysis and processing, fusing of multi-source data is an important research direction in the field of big data science. Aimed to problems that exist in fusing multi-source images under big data science, how to extract features, ringing phenomenon, details information distortion phenomenon and missing low-frequency information problem, this project introduces conditional random field and neural network model to approximate images in order to extract effective information and solve fusion problems based on feature-level image fusion methods. Because of the correlation between the multi-source images, this project uses conditional random field to describe the corresponding region, extract features and can avoid the ringing phenomenon in the fusion results; Then, aimed to details information distortion problem, pulse coupled neural network is introduced to construct a new fusion scheme. And for low-frequency information missing phenomenon, a new fusion strategy based on autoencoder is proposed for fusing low-frequency subbands. Contents of this project follow: according to the characteristics of big data multi-source image, the conditional random field is used to classify and extract the features, and obtain the stable and reliable details from the original images; aimed to reduce the details information distortion phenomenon, the new fusion algorithm is built by combining pulse coupled neural network and traditional fusion strategy; at last, for reserving more low-frequency information that is extracting from original images, a fusion strategy based on deep neural network model is introduced to construct new fusion algorithm and will obtain a new fusion results which are more effective and clear.

作为大数据分析处理的关键环节，多源数据的信息融合是重要的研究方向。本项目针对大数据科学下多源图像融合中存在的问题：大数据多源图像如何提取特征、避免振铃现象、细节信息易损失等，拟在特征级图像融合的基础上，提出基于条件随机场及神经网络模型的融合方法来解决。由于待融合多源图像相对应的像素具有相关性，可采用条件随机场去描述，实现特征融合，可避免在结果中出现振铃现象；引入脉冲耦合神经网络对融合框架进行改进，降低细节信息损失；针对低频信息易缺失，本项目拟引入Autoencoder深度神经网络模型给出低频子带融合策略，构成新算法。项目拟开展的工作包括：根据大数据多源图像本身的特点，采用条件随机场进行特征融合，给出相应的融合算法；以增加融合结果中细节信息为目的，结合脉冲耦合神经网络得到多源图像融合框架；以保存更多有效的低频信息为目的，引入基于深度神经网络模型的低频融合策略，构造有效的多源图像融合方法。

本项目针对大数据分析处理关键环节中的多源遥感数据融合问题，做了几方面的工作：1.从遥感图像数据表示的角度出发，引入机器学习算法，将最小Hausdorff距离与非下采样Shearlet变换相结合，针对传统遥感图像融合方法中存在的保留图像光谱特性和提高空间细节表征能力相冲突的问题，提出新的解决方案；2.从遥感图像特征表征的角度出发，引入稀疏表示理论，构造针对高空间分辨率的字典和低分辨率的字典表征图像特征，并采用K均值聚类对训练样本进行聚类，以提高字典训练的效率，获得自适应字典，表征多光谱图像特征；3.针对大多数融合规则存在单一性的问题，运用多元经验模态分解法将遥感图像分解为多尺度图像，利用稀疏表示对目标图像进行融合，使得目标图像在细节增强的同时光谱信息的损失度尽量减小；4.引入卷积自编码模型，学习多源遥感图像之间的非线性映射关系，从而减少多源遥感图像之间的差异，获取光谱性能和空间分辨率更好的融合图像，为后续的研究提供帮助。