Research On Deblurring And Depth Estimation Of Defocused Images

Currently, in medicine, exploration, archeology and industrial testing, the scenesimages recorded by the endoscopic imaging system are blurred, which is caused byits camera out of focus. The clear scene images and accurate3D visualizationscenes can be achieved by the image deblurring and3D reconstruction from thedefocused images to provide decision-making reference for doctors, explorers,archaeologists and testing personnel. Therefore, the research of image deblurringand depth deblurring from the defocused images has important technical andapplication value.For defocused images, this paper will be developed in both image deblurring anddepth estimation. Image deblurring and depth estimation can be considered assemi-blind deconvolution process and blind deconvolution process respectively.From the perspective of generative model and discriminative model, the blinddeconvolution problems will be transferred to the semi-blind deconvolutionproblems, which can treate image deblurring and depth estimation as semi-blinddeconvolution process. Accordingly, for the image deblurring from defocusedimages, this paper considers in two aspects: the non-negativity constraint iterativemethod of image deblurring and the selection of total variation image deblurringfidelity terms. For the depth estimation based on defocus images, the paperdiscusses about the selection of the criterion function fidelity terms and theregularization terms. In addition, this paper transfers the traditional imagegeneration model into discriminant model and achieves the transformation fromblind deconvolution to sime-blind deconvolution. It also explores the depthestimation based on discrimination technology. Main contents are as follows:From the perspective that the image deblurring belongings to deconvolutionproblem, this paper proposes an iterative algorithm based on Kuhn-Tucker to solvethe non-negative constrained image deblurring problems. The iterative algorithmcan guarantee each iteration function sequence is non-negative, and theoreticallyprove that the sequence converges monotonically to a minimum. Experimentalresults show that the recovered visual effect and PSNR of iterative algorithm based on Kuhn-Tucker are superior to those of the steepest descent method withnon-negative constraint and conjugate gradient method based on projection ontoconvex sets.From the perspective that image deblurring belongs to semi-blind deconvolutionproblem, this paper presents a total variation image deblurring model based onI-divergence. Since the variables involved in image deblurring problems arenon-negative, this model selects I-divergence criterion as fidelity term andcombines total variation regularization term to build functional extreme function,and then converts into Euler-Lagrange equation. Introducing the time variable, itforms gradient descent flow and combines central difference method and windwardprogram structure to construct numerical solution, and then solves the gradientdescent flow. Experimental results show that the recovered visual effect, PSNR andBSNR of total variation image deblurring model based on I-divergence are superiorto those of the total variation image model based on least squares criterion.From the perspective that depth estimation from defocused image belongs toblind deconvolution problem, this paper presents a depth estimation method fromdefocused image method with geometric constraints. The mothod has severaladvantages as follows: Firstly, it chooses the difference degree between thedefocused images as functional extreme function, which is generated by a biggerdegree and a smaller degree of the observed defocused images. Therefore, it canavoid the disadvantage of restoring the image depth information and a clear imageat the same time, and improve the algorithm efficiency; Secondly, based onreal-aperture imaging geometric principles, this method derives a series of intervalconstraints of the relative dispersion parameter in point spread function in differentcircumstances. It effectively avoids low efficiency of the algorithm caused by thecomplex regularization term (such as Nonlocal uniform regularization, TV term,etc.); Thirdly, since the variables involved in depth estimation are non-negativecharacteristics, this method selects the I-divergence criterion as its fidelity term toimprove the algorithm efficiency and the accuracy of depth estimation. simulationresults without noise show that the depth estimation with geometric constraintsmethod is more suitable for continuous smooth surface (cosine surface). in theoperating efficiency and accuracy, depth estimation method with geometric constraints is significantly better than that without geometric constraints.simulation results with noise show that the depth estimation method with geometricconstraints is sensitve to the defocused images with salt and pepper noise andPoisson noise. Real data experimental results show that depth estimation methodwith geometric constraints is significantly better than that without geometricconstraints.From the perspective that depth estimation from defocused image belongs todiscriminant learning techniques problem, this paper presents depth estimation fromdefocused image based on discriminant learning techniques, including discriminantmeasure learning phase and depth estimation phase. In discriminant measurelearning phase, it transfers the depth estimation problem into multi-classificationproblem, and then it determines multi-classification discriminant function forms.In order to learn discriminant measure of each depth, the criterion function isconstructed. The criterion function fully considers the distance within the group andthe distance between the two groups. it finally is converted into semi-definiteprogramming using projection sub-gradient descent algorithm to solve them andobtain discriminant function of each depth. In depth estimation phase, for each pixel,calculating N discriminant function values, the pixel depth is classified into theminimum discriminant function value depth. Depth estimation method based ondiscrimination measure can restore the real image depth information learning fromthe simulation image discriminant measure, which is one advantage of the method.As involved operations are all simple matrix operations, and each pixel opposites,this method has strong parallel processing capability and high efficiency. Simulationand real data experimental results show that, in the RMS aspect, the depthestimation method based on the discriminant measure learning is superior to thedepth estimation methods based on singular value decomposition learning.