Research On Super-resolution Reconstruction Algorithm For Normal And Infrared Images

Since the generative adversarial networks have been introduced to image superresolution reconstruction,more and more related models have been proposed to improve the visual quality of super-resolution reconstructed images and image quality.How to design the appropriate algorithmic models for different types of images and make the algorithms lighter so that they can be applied practically is a worthy topic for further exploration and study.In this thesis,we design the adversarial generative network models adapted to the features of visible and infrared images in different practical environments to improve the subjective visual effect while reducing the model complexity of the algorithm and avoiding the model collapse and gradient disappearance.The main research contents include:(1)To address the image quality problems such as blurred edges and insufficient texture details in the learning-based image super-resolution reconstruction algorithm,and the large size of the total number of network parameters and high complexity of the algorithm,we propose a super-resolution reconstruction algorithm based on the depthwise separable convolutional residual module combined with adversarial generative networks: depthwise separable convolution dense block-generative adversarial networks(DSCSRGAN).First,a new depthwise separable convolutional residual block is designed as the base building block for the generator network,which improves the learning and extraction of image features while significantly reducing the total number of parameters,and for the discriminator network,the batch normalization layer is removed to reduce the image artifacts.Second,the frequency-energy similarity loss function is designed to constrain the generator network to generate better quality super-resolution reconstruction images.(2)A new super-resolution reconstruction algorithm: heterogeneous convolutionalwasserstein gan(Het SRWGAN)based on heterogeneous kernel convolution combinedwith the generative adversarial networks is designed to address the problems of IR imageswith simple information,single texture,few features,and single pattern.Firstly,a plug-and-play heterogeneous kernel convolution is introduced,in which a plug-and-playheterogeneous kernel convolution residual block is designed.The introduction of thisresidual block can increase the perceptual field of the neural network and obtain richerimage detail texture and structure information compared with the traditional convolutionmodule.Secondly,it can also reduce the number of model parameters to avoid the modelcollapse occurring,and at the same time reduce the training time and the model's requirement for training arithmetic power,expanding the application field of the algorithm.To better implement supervised learning,a gradient loss function for super-resolution reconstruction of infrared images is proposed to restrain the generator network.(3)Inspired by transfer learning,We propose a method that can use visible images and infrared images with few samples as training datasets for IR image super-resolution reconstruction: progressive super-resolution generative adversarial network(PSRGAN)and an adaptive multi-stage migration learning strategy.By using an information distillation method to obtain the feature distribution of visible images,the PSRGAN is combined with a multi-stage migration learning strategy to further map the abstract features of infrared images and visible images into the same feature space for super-resolution reconstruction of infrared images.In the experiment,PSRGAN combined with the multistage transfer learning strategy can fully utilize two different types of images to achieve super-resolution reconstruction of infrared images,and achieve good results in both subjective visual effects and objective index evaluation.