Research On Infrared Single Pixel Imaging Based On Generative Adversarial Networks

Video images are one of the main forms in modern media.They are indispensable in television,movies,internet videos and surveillance systems.High resolution images can provide more detailed information which is crucial for many applications.Especially in the field of computer vision,however,the progress of computer vision is limited by high-quality images.And obtaining high-quality images has always been a challenge in computer vision research.Generative adversarial networks have attracted widespread attention as a new type of deep learning method.It can generate high-quality and realistic images,further promoting the development of image super-resolution reconstruction towards high-quality direction.For infrared single pixel imaging systems,there is always a contradiction between the sensitivity and resolution of the seeker.This thesis aims to improve the resolution of the image as much as possible in the highly sensitive infrared rose line point scanning method,and to fill in missing information that has not been scanned.The main work is to use rose line scanning for physical compression imaging.Replacing traditional optical reflection systems with optical lenses can effectively reduce losses in optical transmission,while combining deep learning neural networks for control.Through an improved generative adversarial network,an infrared single pixel imaging system that combines sparse sampling and recovery algorithms is trained.Experiments on infrared aerial target datasets have shown that when the input is sparse images sampled with rose lines,single pixel restoration imaging of infrared images is ultimately achieved,ensuring high sensitivity while improving image resolution.Firstly,this thesis describes the single pixel imaging model and image super-resolution reconstruction method and analyzes their principles and derivation process.And then,it introduces the principle and related mathematical models of generative adversarial networks.And based on the characteristics of generative adversarial networks that can present clearer map details,deep learning is applied to super-resolution reconstruction of infrared single pixel images.This thesis proposes an infrared single pixel imaging method based on generative adversarial networks.Firstly,the structure of the deep convolutional network designed in this thesis and how to perform rose line scanning imaging are described.Then,the datasets and parameter configuration used are introduced.Finally,by analyzing experimental simulation results,it is not only verified that this method is effective compared to traditional algorithms,but also demonstrates that the established single pixel imaging model can effectively reconstruct missing images.Secondly,this thesis proposes an infrared single pixel imaging method based on multi feature fusion.By combining wavelet domain with multi-scale constraints on the generator.UNet network is used to improve the discriminator.And attention mechanism is used to determine the authenticity of the image while combining the image features before downsampling.Multiple features are used for more complete feature extraction and fitting.After providing the flowchart of the method and the structure of the network,simulation experiments were conducted on it.Finally,after analyzing the simulation results,better reconstruction results are achieved under the same data and conditions.This further demonstrates the advantages of deep learning algorithms and the practical applicability of this single pixel imaging model.