遥感图像的配准和拼接是遥感图像处理的重要组成部分，已广泛应用在国防军事安全、自然灾害分析、农林资源考察、城市合理规划等多个领域。然而，由于遥感图像复杂的成像机制和成像条件，传统的图像配准算法提取特征繁琐，容易出现特征点误匹配的情况，导致图像配准精度较低，直接影响了遥感图像的拼接质量。因此，论文从空间参数的预测和最佳缝合线两个角度出发，深入研究了遥感图像的配准和拼接问题，给出了基于卷积神经网络和校正网络的遥感图像配准算法。与此同时，将基于视觉感知的能量函数缝合线算法应用于图像拼接处理中。
论文将卷积神经网络和校正网络相结合的遥感图像配准算法运用于遥感图像的配准过程中，利用卷积神经网络直接预测图像之间的变换参数，再结合校正网络寻求最佳的变换参数。首先，算法采用仿射变换网络对参考图像进行仿射变换，批量生成可用训练样本。其次，算法将特征提取和特征匹配放在卷积神经网络端到端的架构中，自学习参考图像及浮动图像之间的形变模型，直接预测图像之间的仿射变换参数。最后，算法采用校正网络反向传播对卷积神经网络的预测结果进行校正，实现遥感图像更加精确的配准。实验结果表明，相比于常用的SIFT算法、SURF算法和基于深度神经网络的配准算法，论文的配准算法在图像的配准精度和速度上，均有显著提升。
为了使遥感图像有一个较好的拼接后视觉效果，论文采用了基于视觉感知的能量函数缝合线算法将人类视觉的非线性和显著性转化为能量值最小的函数。算法通过Sigmoid函数模拟视觉色差的感知范围，调整显著性目标权重，得到待拼接图像间的视觉最佳缝合线，最后采用多区域渐入渐出加权平均法进行图像融合拼接处理。实验结果表明，相比于基于色差缝合线算法和基于颜色梯度缝合线算法，论文算法的图像拼接结果信息更加丰富，直观视觉效果更佳。

Remote sensing image registration and stitching technology are important parts of remote sensing image processing. It has been widely used in many fields such as national defense military security, natural disaster analysis, agricultural and forestry resources investigation, urban rational planning and so on. However, due to the complex imaging mechanism and imaging conditions of remote sensing images, the traditional image registration algorithms are cumbersome to extract features and prone to mismatch the feature points, which directly affects the stitching quality of remote sensing images. Therefore, from the perspective of spatial parameter prediction and optimal stitching, the thesis deeply studies the registration and splicing of remote sensing images, and then gives a remote sensing image registration algorithm based on convolutional neural network and correction network. At the same time, the energy function stitching algorithm based on visual perception is applied to the image stitching process.
In the thesis, a remote sensing image registration algorithm combining convolutional neural network and correction network is used in the registration process of remote sensing images. The convolutional neural network is used to directly predict the transformation parameters between images, and then combined with the correction network to find the best transformation parameter. First, the algorithm uses an affine transformation network to affine transform the reference image to generate training samples in batches. Secondly, the algorithm puts feature extraction and feature matching in the end-to-end architecture of the convolutional neural network, learns the deformation model between the reference image and the floating image, and directly predicts the affine transformation parameters between the images. Finally, the algorithm uses correction network to correct the prediction results of the convolutional neural network to achieve more accurate registration of remote sensing images. Experimental results show that, compared with the commonly used SIFT algorithm, SURF algorithm and deep neural network algorithm, the algorithm in this thesis has significantly improved the accuracy and speed of image registration.
In order to make the remote-sensing image have a better visual effect after stitching, the paper uses an energy function stitching algorithm based on visual perception to convert the non-linearity and saliency of human vision into a function with the smallest energy value. The sigmoid function is used to simulate the perception of visual chromatic aberration and the weight of the saliency target is adjusted to obtain the visual most between the images to be stitched. Finally algorithm in the thesis uses the weighted average method of multi-region gradual in and out to carry out image fusion stitching. The experimental results show that the image stitching results of the algorithm in the thesis is more informative and visually better than the color difference stitching algorithm and the color gradient stitching algorithm.