Research On The Key Technology Of Image Motion Deblurring For Dynamic Visual Inspection

Dynamic visual inspection generates motion blurred images and unreliable results.Existing approaches of image motion deblurring has relatively low performance for estimation,and cannot handle outliers properly,moreover,they also suffer from time-consuming optimization process.This research has focused on the key technology of image motion deblurring for dynamic visual inspection,specifically aimed to address the issue for motion blur kernel estimation,deconvolution modeling and ringing artifacts suppressing,which contributes to facilitate the intelligentization and the automation of inspection equipment.This research was funded by the Guangzhou Science and Technology Plan Project(201802030006)and the Open Project Program of Guangdong Key Laboratory of Modern Geometry and Mechanics Metrology Technology(SCMKF201801).This research aims to deblur the motion blurred image for dynamic visual inspection,and it has reviewed the recent progress on motion blur kernel estimation,deconvolution modeling and ringing artifacts suppressing.The main contributions are unfolded as follows:(i)The pipeline of dynamic visual inspection was analyzed here to design the framework for image motion deblurring,and the key technologies of feature extraction-based motion blur kernel estimation,deep learning-based deconvolution modeling and ringing artifacts suppressing for dynamic visual inspection have been paid close attention to,along with their implementation schemes.(ii)According to the analysis on linear motion blur kernel estimation,the relation between the orientation of stripes in spectrum and blur angle was inferred to estimate the blur angle,while the distribution of autocorrelation function on salient edge was exploited to estimate the blur length,yielding the linear motion blur kernel estimation for arbitrary orientation;To utilize the orientation information of key points of autocorrelation function on salient edge,the function was visualized as a 3D surface.Then the relation between the orientation/positions and the two parameters of the blur kernel was obtained,leaving the two parameters of the linear motion blur kernel separately estimated;Experimental results have shown that our method still performed well at SNR = 20 d B noise level.(iii)According to the analysis on deconvolution modeling,the scheme of deep learningbased deconvolution modeling for dynamic visual inspection was proposed on the characteristics of deconvolution;The energy function of deconvolution was decoupled as data fidelity and regularization sub-problems via variable splitting method,and the task-driven fidelity sub-problem was respectively modelled based on Gaussian and Poission distributions to guarantee the goodness-of-fit;The data-driven regularization sub-problem has incorporated CNN(Convolutional Neural Network)to adaptively learn the image prior at a wide range of noise level,where a hand-crafted regularizer was no longer needed.The proposed method delivered a fast testing speed and can be transferrd across different vision tasks.The CNN learned prior has served as a plug-and-play part,giving rise to the integration of model-based and learning-based deconvolutions;Experimental results have shown that our method has improved PSNR by an average of 30.79%,up to 31.75 d B,and achieved two orders of magnitude faster speed on average than the conventional approaches.(iv)According to the analysis on ringing artifacts suppressing,the scheme of ringing artifacts suppressing for dynamic visual inspection was proposed on the mechanism of the clipped boundary and the saturated pixels;The motion blurred image was extrapolated to generate a tile where the discontinuities of clipped boundary were reduced;The Gaussian and the Poission based fidelity sub-problems were respectively modified and improved to alternate with FFDNet(Fast and Flexible Denoising Network)to reject the saturated pixels in deconvolution,yielding better visual quality in deblurred image and the full version of image motion deblurring for dynamic visual inspection;Experimental results have shown that compared with existing approaches,our method had better SSIM and FSIM increasing by 0.05%—20.42%,and exhibited superior performance in visual quality and run time.(v)The application platform was set up with its software and hardware.Then the image motion deblurring for dynamic visual inspection was subjected to object and size detection in dynamic scenes.The improvement in accuracy helped further assess the deblurred visual quality,while the feasibility and the effectiveness of the image motion deblurring for dynamic visual inspection was better demonstrated;Experimental results have shown that using our method,the accuracy of object detection has increased by an average of 35.48%,and the error of size detection has decreased by an average of 98.23%.