Research On Infrared Dim-small Target Super-resolution Restoration Arithmetic Based On POCS

With the spring up of the infrared imaging related industry, the infrared imagingtechnology has become the mainstream development direction of the intelligentphotoelectrical detection due to its good concealment, wide detection range, highpositioning accuracy, long distant penetration, light weight, little volume, low powerdissipation and high solidity. However, the features of the image of infrareddim-small target such as less details and low SNR become the bottleneck of theapplication of infrared image. How to enhance the imaging effect of the infrareddim-small target becomes the hotspot of the research. Starting from the point of“restoration as foundation”, the theory and technology of the infrared dim-smalltarget super-resolution restoration by utilizing the theory and technology of thesuper-resolution restoration are explored in this thesis.This thesis mainly focuses on the research of super-resolution restorationalgorithms of the infrared dim-small target based on POCS. Aiming at solving thesuper-resolution restoration problem of the infrared dim-small target, the traditionalsuper-resolution restoration algorithm of POCS is optimized. And four improvedalgorithms are proposed which improved the performance. Meanwhile, thealgorithms are realized in real-time or near real-time which can be applied in thepractical infrared image processing system. This thesis proposes four improved POCS algorithms and a new evaluationmethod of the super-resolution restoration. And the effectiveness of the improvedalgorithms and the evaluation method are evaluated by the infrared dynamic scenesimulation system and the infrared image processing system.The main work and innovation of this thesis are:(1) For the noise sensitive problem of the traditional POCS restorationalgorithm, the BM3D filtering method with better de-noising effect and the POCSrestoration algorithm are combined in this thesis. We optimize the BM3D methodand propose the method of mean pre-screened image block and limiting the numberof packet image blocks to reduce the computation of BM3D method. Experimentalresults show that the proposed POCS based on BM3D can achieve better restorationeffect than that of the traditional POCS method when the low resolution imagecontains noise, furthermore no noise in the high resolution image can be perceivedbasically.(2) For the disadvantage of the traditional super-resolution restorationevaluation system only concerning about a particular aspect of the statisticalproperties of the image, we propose the super-resolution restoration evaluationmethod based on SSIM_NCCDFT, which combines the gray value and contrast ofthe spatial domain and the autocorrelation of frequency domain. Therefore, theproposed evaluation method can evaluate the results of the super-resolutionrestoration in both spatial domain and frequency domain. Experimtnal results showthat the evaluation method can well evaluate the super-resolution restoration results.Furthermore this evaluation method has some significance for super-resolutionrestoration evaluation(3) For the long iteration of the POCS super-resolution restoration algorithmand the shortcomings of incapability to meet the real-time detecting of opticaldetection system, we propose a fast POCS super-resolution restoration algorithmbased on the gradient image, which classifies image pixel according to the gradientof the image, and then uses different iteration factor to calculate. The iteration step is larger when the gradient is bigger and the iteration step is smaller when the gradientis smaller. The improved algorithm can preserve edge information and suppressnoise. Therefore, it can guarantee the performance of the super-resolution restorationand greatly reduce the running time. Simultaneously, another fast POCSsuper-resolution restoration algorithm based on region selection is proposed. Thetarget area is the key point we focus on in the optical detection system, while thisarea contains only very small number of pixels. Therefore, we use thresholdsegmentation and combination to acquire the union of all target areas. Then weexecute super-resolution restoration only in the union of all target areas. In this waywe decrease the huge computation of background restoration and greatly reduce theoperation time to achieve real-time or near real-time. So this super-resolutionrestoration algorithm can be applied in the practical infrared image processingsystem.