Research On Super-resolution Reconstruction Of Infrared Imaging System

The spatial resolution of image is an important technical index of infrared imaging system,which directly determines the application prospect of imaging system in the fields of remote sensing infrared imaging,fault detection,medical image analysis and foe recognition.However,the current infrared imaging technology still has some shortcomings,such as low signal-to-noise ratio,low contrast and less high frequency detail information.How to improve the resolution of the image obtained by the infrared imaging system has become an important problem that people need to solve urgently.The scheme to improve the hardware design such as increasing the size of the photoreceptor and reducing the pixel size is the most direct way to improve the imaging resolution.However,these methods are costly and have an unbreakable physical limit,which makes them only limited to improve the resolution of the imaging system.Therefore,the thesis will focus on the super-resolution reconstruction technology of infrared imaging system.In this paper,we propose three super-resolution reconstruction algorithms and discuss a verification method of super-resolution reconstruction with joint image registration,and finally design a real-time infrared super-resolution imaging system.The main work and innovation of this paper include:(1)Single image super-resolution reconstruction.The thesis first studies three kinds of polynomial interpolation super-resolution reconstruction techniques and two learning-based super-resolution reconstruction techniques.Aiming at the problem that the high resolution image obtained by the generative adversarial super-resolution reconstructed network only has highly realistic visual effect,but the objective quality evaluation index is not high,this paper proposes a super-resolution reconstruction technique of fusion network.Reconstruction network consists of three parts,of which the generative adversarial network reconstructs the initial high-resolution images,the gradient transformation network obtains high-frequency details of the imageinformation,and the fusion network fusion these two image information to obtain the final high resolution image.The fusion reconstruction network can effectively introduce external sample knowledge.The reconstructed image has real visual effects and clear details.Compared with the comparison algorithm,the objective evaluation index is optimal.(2)Multi-frame image super-resolution reconstruction.The main difference between multi-frame image reconstruction and single-frame image reconstruction is that multi-frame reconstruction introduces the time-space correlation between the image sequences,thereby incorporating different high-frequency detail information between the input images.In the reconstruction constraint super-resolution reconstruction,the image registration and reconstruction have the relationship of mutual promotion and mutual restriction.On this basis,a super resolution reconstruction method based on joint image registration is proposed,that is,loop iteration is adopted during reconstruction,and image registration parameters and reconstruction effects are continuously corrected until the final high-resolution image is obtained by iterative convergence.Compared with the traditional reconstruction algorithm,the high-resolution image of the joint reconstruction algorithm obtains a sharper edge and more detailed performance.The constraint reconstruction method can effectively use the different information between the sequence images,but the external knowledge that can be used in the reconstruction is less.Therefore,the paper proposes a multi-frame image super-resolution reconstruction technique based on generative adversarial network,which expands the generative adversarial super-resolution reconstruction network to multi-frame by adding image registration and weight representation layer.The generative adversarial multi-frame reconstruction algorithm can not only fuse different information between sequence images,but also can introduce external sample details.Its reconstructed image has more complete high-frequency detail information and real and natural image effect,and has a strong advantage in the objective evaluation index.(3)A verification method of super-resolution reconstruction based on imageregistration.In order to verify the effectiveness of super resolution reconstruction technology in improving image resolution,a new verification method based on image registration for super-resolution reconstruction is proposed in this paper.The registration experiments and noise experiments show that the super-resolution reconstruction technique can not only improve the image registration accuracy,but also improve the anti-interference ability of the registration noise,which shows that the joint verification method is consistent with the evaluation results of the subjective and objective evaluation methods,and has a stronger expression for the accuracy of the image super-resolution reconstruction results.(4)An infrared super-resolution imaging system based on controllable micro-displacement.Single-frame image reconstruction and multi-frame image reconstruction are mostly image post-processing,and their computational complexity is usually high,which often can not meet the real-time requirements of the application of super-resolution reconstruction.Therefore,our paper designs a super-resolution imaging system based on controlled micro-displacement.The imaging system controls the displacement of the rear lens of the infrared lens through the controllable micro-displacement platform so as to obtain the infrared image sequence with sub-pixel displacement known.Finally,the image sequence is reconstructed by using the non-uniform interpolation method,so as obtain high-resolution images in real time.