Research On Deblurring Algorithm For Vehicular Image

Vehicle Navigation System based on the combination of Inertial Positioning andMachine Vision (IP/Vision) is a new research area. In IP/Vision system, InertialPositioning plays an important part in Vehicle Navigation System, but it suffers fromcumulative error which is relatively large for precise vehicle positioning. By MachineVision the vehicle location can be pinpointed to a very small coverage of road marksobserved by MV, thus performance of INS is improved. In Machine Vision signpostsand road markers are captured by the on-board image/video devices, such as vehicularcamera, then by image processing tools the markers are identified and to be used tomatch the signs/marks in digital map to correct the accumulated error. However, theobtained images are usually blurred due to the vehicular camera shake or motion, whichdegrades identification performance. Therefore, this thesis is focus on how to deblursuch kind of degraded images. On basis of the currently proposed image deblurringalgorithms and taking the characteristics of vehicular blurry image into account, severalexcellent deblurring algorithms are improved to efficiently restore sharp vehicularimages. The three main works of this thesis are described as follows:First, according to the problem of the two classical image non-blind deblurringalgorithms, Fast Image Deblurring using Total Variation (FID-TV) and Fast ImageDeblurring using Hyper-Laplacian Priors (FID-HLP), a more robust and fasteralgorithm is proposed to deblur vehicular blurry images (FID-ADM: Fast VehicularImage Deblurring based on Alternating Direction Algorithms), which gives a noveldeblurring model that can be optimized by Alternating Direction Algorithms. Numerouscomparative experimental results demonstrate that the proposed method performsfavorably against FID-HLP on both speed and image quality for moderate sized blurryimage; and for the large-scale blurry images, it has obvious advantages on image qualitythough its speed is slower than FID-HLP.Second, Blind Vehicular Image Deblurring uses the Normalized Sparsity Measure(BID-NSM) and its weaknesses are reviewed briefly. We adopt the idea of kernelestimation using Normalized Sparsity Measure, and then recover sharp image with previously proposed non-blind deblurring methods (FID-ADM). Lots of convictiveexperiments show that the improved method is superior to BID-NSM with respect to thespeed and quality of restored images.Finally, considering that the efficiency and accuracy issues of Normalized SparsityMeasure kernel estimation method, when kernel size is larger than object size, whichwill result in inaccurate kernel and deblured image with severe ringing artifacts. Hence,a blind image deblurring based on the idea of kernel refinement is proposed to recoversharp vehicular images. To suppress possible noise and small details that can disturbkernel estimation, a bilateral filter is firstly applied to preprocess input blurry image,and then the true kernel is obtained by the process of kernel refinement method, finally,the previously proposed non-blind debluring method (FID-ADM) is used to recoversharp image with the refined true kernel. Through the experiments, the proposedapproach has been demonstrated that: it can rule out ringing artifacts efficiently; itsestimated kernel can approximate true kernel effectively. The recovered image can meetthe practical application.