Super-resolution of large field of view infrared image based on residual convolutional auto-encoders

Abstract

Because of its imaging characteristics, large field of view infrared images have low resolution and few high-frequency details. And it is difficult to obtain high-resolution large field of view infrared images as training sample database. Therefore, some super-resolution reconstruction algorithms that achieve better results in visible images may not be suitable for large field of view infrared images. Based on Convolutional Auto-encoders, a shallow Convolutional Auto-encoders which consists of four convolution layers, one maxpooling layer and one upsampling layer was constructed. And deep Convolutional Auto-encoders composed of 10 convolution layers, 2 maxpooling layers and 2 upsampling layers was provided. shallow Residual Convolutional Auto-encoders and deep Residual Convolutional Auto-encoders also were concluded. In order to make up for the shortcomings of large field of view infrared images, this paper adopts the combination of some large field of view infrared images and ordinary field of view infrared images as the image training library, and uses the large field of view infrared image data as the test set. The reconstructed image quality and the objective indexes of MSE, PNSR and SSIM show that the reconstruction effect of the shallow Residual Convolutional Auto-encoders is slightly better than that of the deep Residual Convolutional Auto-encoders, while the shallow s Convolutional Auto-encoders and the deep Convolutional Auto-encoders without residual model are not good. However, the reconstruction effect of deep Convolutional Auto-encoders is the worst due to the loss of more shallow information.