Research On Super-Resolution Reconstruction Method Of Remote Sensing Images

With the development of remote sensing technology,high-quality remote sensing images have been widely used in meteorological observation,disaster monitoring,geographical mapping,military surveillance,land resources mapping and remote sensing target detection.However,the quality of remote sensing images obtained is low due to factors such as long-distance,complex targets,and insufficient imaging system resolution.The emergence of super-resolution reconstruction technology breaks through the limitations of the hardware system,can improve the resolution of images,restore the details of images,and bring a new direction to the quality improvement of remote sensing images.Nowadays,deep learning has rapidly promoted the development of super-resolution reconstruction technology,and reconstruction methods based on convolutional neural network have emerged continuously and played a vital role in image reconstruction,so the image quality has been effectively improved.Considering the complex problems in the process of remote sensing image reconstruction,two kinds of remote sensing image super-resolution reconstruction algorithms based on convolutional neural network are proposed.The specific research contents are as follows:(1)Aiming at a large number of operations in CNN method and the model’s dependence on convolution operation,a super-resolution reconstruction network model with multi-scale feature self-attention mechanism is proposed.This model first uses different convolution kernels to extract features from images to obtain feature information of different scales.Then,a self-attention fusion module composed of spatial attention and channel attention in parallel is used to establish the relationship between different areas of the image.Proper channels are selected to adapt to complex remote sensing images.Different levels of information obtained are retained as much as possible,hence improving its ability to represent and learn global features.The image is reconstructed by adaptive after upsamling is performed on feature information.During the whole reconstruction process,the simple structure of the model reduces unnecessary procedures and optimizes the number of network layers.The experimental results show that the reconstructed image positively affects both subjective and objective evaluation.Especially when the amplification factor is 2,the PSNR value of this algorithm is 0.16-4.11 d B higher than other algorithms,and the SSIM value reaches 0.9602.Also,the reconstructed image is more precise.(2)Aiming at the problem of oversimplification and loss of details in the feature extraction of complex remote sensing images,an improved convolutional neural network super-resolution reconstruction method is proposed.Combining multi-scale feature extraction with residual channel attention,three different kernels are used first to extract multi-scale superficial features from low-resolution images.Then,six residual channel attention blocks are used to obtain in-depth,informative features.Meanwhile,the same recursive layer is used for feature processing,and the receptive field gradually increases in the recursive process while enhancing the context information.By using long-short skip connections to fuse different features in the whole network,the inefficiency of feature information in the transmission process is avoided.Deep CNN is used to reduce the amount of computation further,and a parallel 1×1 network is added to reduce the input dimension to speed up image information processing.Finally,better reconstruction effect can be achieved by integrating high and low frequency information during reconstruction.Experimental results show that the proposed network model achieves better results than other models.When the amplification factor is 2,the PSNR value is 0.32-4.38 d B higher than different algorithms,and the SSIM value reaches 0.9612.The reconstructed image extracts more features and is richer in details.