Super-resolution Reconstruction of Night-light Images Based on Improved SRCNN

Night-light remote sensing images have been used in numerous fields and have brought much value to human society. However, the low resolution is a major drawback of night-light images, limiting its application in many ways. Image processing research proposes super-resolution reconstruction technology, which can obtain high-resolution images from low-resolution images. It is often used in some fields that need image details. We improve the SRCNN network structure by removing the bicubic interpolation with high computational complexity and replacing it with a deconvolutional layer. We added a feature extraction layer. We can extract more features in the image by using two convolution layers to extract the enlarged image. Finally, we also optimize the convolutional layer of the nonlinear mapping to reduce the training time. We produced a training set and a test set of night light images to evaluate the effect of network improvement. Through the final comparison, our method can improve the resolution of luminous image, which is greatly improved compared with SRCNN.