| Remote sensing image contains a lot of ground information,which has important application value in military and civil fields,such as military target strike,land monitoring,vegetation coverage detection and forest fire monitoring.However,due to the complexity of remote sensing imaging process,it is not easy to obtain high-resolution remote sensing images.Remote sensing image super-resolution reconstruction technology is a kind of technology that breaks through the limitation of imaging system and generates high-resolution image from low-resolution image by software processing method under the condition of existing remote sensing imaging technology.In this paper,the advanced convolutional neural network super-resolution is studied and improved,and the specific research content and innovation are as follows:1.An improved algorithm for remote sensing image super-resolution reconstruction using residual network is proposed,which is improved on the basis of VDSR algorithm.In the improved network,the number of the convolution layer increases to 24.The activation function is changed from ReLU function to PReLU function.L2 norm is used as the loss function in network training,and the optimizer is changed from mini-batch gradient descent to Adam algorithm.In the process of training and testing,the image data set in the network is also changed to practical remote sensing images.The network training experiment and network generalization experiment are carried out for the improved network,and the experimental results and data are compared and analyzed.2.An improved algorithm for remote sensing image super-resolution reconstruction using densely connected network is proposed,which combines our improved SR residual network and densely connected network,and the residual blocks of our improved SR residual network are replaced by dense blocks.In the structure of our improved SR residual network,one residual block is improved to two groups of dense parts,and each group of dense part is made up of 8 dense blocks.At the same time,the deconvolutional layer is added to the network in order to enlarge the image size directly.The loss function is changed from L2 norm to L1 norm.The improved network is trained and tested by remote sensing images.The network training experiments and network generalization experiments are carried out,and the results are compared and analyzed.3.The software platform of super-resolution of remote sensing image is developed based on Qt5 interface library.The software can display the effect of a variety of remote sensing images super-resolution in a friendly way,which makes it more convenient for users to run every super-resolution algorithm,and visually compare the results of remote sensing image super-resolution reconstruction. |
