密集卷积残差网络的遥感图像融合

针对传统的遥感图像融合方法通常会引起光谱失真的问题和大多数基于深度学习的融合方法忽略充分利用每个卷积层信息的不足,本文结合密集连接卷积网络和残差网络的特性,提出了一个新的融合网络.该网络通过建立多个密集卷积块来充分利用卷积层的分级特征,同时块与块之间通过过渡层加快信息流动,从而最大程度地对特征进行极致利用并提取到丰富的...

Pan-sharpening by residual network with dense convolution for remote sensing images

Pan-sharpening (also known as remote sensing image fusion) aims to generate Multi-Spectral (MS) images with high spatial resolution and high spectral resolution by fusing high spatial resolution panchromatic (PAN) images and high spectral resolution MS images with low spatial resolution. Traditional pan-sharpening methods mainly include the component substitution method, the multiresolution analysis method, and the model-based optimization method. These fusion methods involve linear models, which are difficult to use in achieving the appropriate trade-off between spatial improvement and spectral preservation. In addition, they often introduce spectral or spatial distortion. Recently, many fusion methods based on deep learning have been proposed. However, their network depth is relatively shallow, and detailed information is inevitably lost during feature transfer. Hence, we propose a deep residual network with dense convolution for pan-sharpening.As the network becomes deep, the features of different levels become complementary to one other. However, most fusion methods based on deep learning ignore making full use of the information of each convolution layer. The densely connected convolutional network allows the features of all previous layers to be used as input for each layer in one densely connected block. To fully utilize the features learned from all convolution layers, we establish the multiple densely convolutional blocks to reuse features. Moreover, the information flow is accelerated by the transition layer between every two blocks. These maximize the use of features and extract rich features. Given the great correlation between deep features and shallow features, residual learning is used to supervise the densely convolutional structure to learn the difference between them, that is, residual features. Thus, residual learning combines shallow features and residual features to obtain further advanced information from MS and PAN images, which prepares for obtaining fused images with high spatial and spectral resolution.To evaluate the effectiveness of the proposed method, we conduct simulated and real-image experiments on the 4-band GaoFen-1 data and 8-band WorldView-2 data with multiple land types. The trained network is generalized well to WorldView-3 images without pre-training. The visual and the quantitative assessment results show that the high-resolution fused images obtained by using the proposed method are superior to the results produced by the traditional and deep learning methods. The proposed approach achieves high spectral fidelity and enhances spatial details by reusing features.The proposed method makes comprehensive use of the advantages of densely convolutional blocks and residual learning. In the feature extraction stage, different levels of features are connected in series through the densely convolutional blocks. This characteristic makes the transmission of features and gradients effective in alleviating the gradient disappearance problem and provides rich spatial and spectral feature for fusion results. In the feature fusion stage, residual learning is used to learn the difference between deep features and shallow features, that is, residual feature. Hence, the convergence speed of the network is accelerated. The experiment result shows that our network has good fusion and generalization abilities.