Airborne Rotating Video Image Stabilization And Deblurring Technology

Airborne photoelectric detection is an important means of aerial surveillance,reconnaissance and attack.The motion and vibration of airborne aircraft,the influence of atmospheric turbulence and other factors are easy to cause the motion blur of airborne video,affect the effect of aerial reconnaissance,and reduce the ability of target detection and feature fine recognition.In this paper,the real-time image stabilization technology of airborne video is studied,and the off-line processing technology for airborne video blur is studied.First of all,an adaptive real-time image stabilization algorithm based on Shi-Tomasi feature points^[1]is proposed for the interframe jitter of airborne video caused by the motion of the carrier and the rotation of the photoelectric platform.To solve the problem of rotation image stabilization in airborne video due to sparse and concentrated feature points extracted from sky,cloud and fuzzy background,an adaptive threshold Shi-Tomasi corner detection algorithm was adopted to improve the detection ability and distribution uniformity of feature points.In view of the difference of rotation jitter and translation jitter components,a hybrid Kalman filter algorithm is used to improve the robustness of complex jitter image stabilization.The test results show that the proposed algorithm has good image stabilization ability for airborne rotating video jitter within10 degrees,and can meet the real-time processing requirements.To solve the problem of airborne video blurring caused by atmospheric turbulence and fast motion,a neural network algorithm of multi-frame pre-alignment is proposed in this paper.The target frame is intercepted by target tracking technology for pre-alignment,so as to solve the problem of multi-frame automatic alignment of neural network due to large range of target motion.The combined convolution module is used to replace the deformation convolution module in EDVR to improve the stability and real-time performance.Meanwhile,the AAD airborne superresolution dataset was constructed for network training to improve the adaptability of the network to air scenes.Comparison test results show that the superresolution effect of the proposed algorithm is significantly better than that of other classical algorithms on public datasets,and the training stability is improved without loss of performance.Good deblurring effect has also been achieved in practical engineering,which greatly improves the ability of identifying fine features of aerial targets.