Weighted Aggregation Guided Filtering And Application Research

With the emergence of digital images,the edge preserving filtering technology has been proposed and applied.After decades of development,the edge preserving filtering technology has been widely applied in image fusion,edge detection,dehazing and other classic fields.Therefore,the edge-preserving filtering technology plays an increasingly important role in computer vision and other fields.Compared with the bilateral filtering,the guided image filtering has lower algorithm complexity(O(N))and can still achieve the same effect as the bilateral filter.As a local filtering algorithm,guided image filtering still has the common shortcomings of local algorithms that it may make some weak image edges blurred in the final output image while smoothing the image and halo artifacts may be introduced into the output image.The works of this thesis can be summarized as follows:1.Firstly,based on the principle of guided filtering,the aggregation strategy of guided image filtering is analyzed,and then a novel weighted aggregation strategy is proposed and the guided image filtering is developed.By applying the weighted aggregation strategy,the developed guided image filtering(WAGIF)can obtain more sharp edges than the original guided filtering while the input image is smoothed,therefore halo artifacts are greatly suppressed in the filtered image.2.In order to solve the contradiction between filtering out error disparities and maintaining disparity edges in the traditional disparity optimization methods,a new linear weighted median filtering based on the proposed guided image filtering with weighted aggregation is proposed and applied to the stereo disparity optimization stage.Considering the outperformance of the proposed WAGIF in edges preserving,in this thesis,a weighted median filtering is introduced by combining with the local histogram technology and WGIF.More specifically,the proposed weighted median filtering has a constant computational complexity which can be applied in some real-time applications.Then,the weighted median filter is applied to the stereo disparity optimization.Finally,the experimental results show that,with the help of the weighted median filtering,edge features in disparity maps are preserved while matching errors are significantly removed.Even if it is applied to some early and disadvantage stereo matching algorithms,the final disparity maps also have high accuracy.3.In the following part,the proposed WAGIF algorithm is also introduced into the multi-exposure high dynamic image fusion and a full details enhanced exposure fusion algorithm is proposed.In this exposure fusion framework,the brightness component of images having different exposure levels are processed by the WAGIF to decompose detail layers;then the decomposed detail layers are fused by Laplace pyramid fusion algorithm to get the final fused images which have fine details.The experimental results show that the proposed algorithm has a better edge-preserving effect than other algorithms,and the fusion result is natural,subjectively.Furthermore,the proposed algorithm is better than other fusion algorithms in several object evaluation metrics,such as information entropy,mutual information and average gradient.4.At last,to address the problem of halo in the final defogging image caused by the guided filter in the traditional dark channel prior defogging algorithm,the proposed WAGIF algorithm is applied to the transmittance optimization step in the dark channel prior defogging algorithm framework.Compared with other defogging algorithms,the proposed optimized transmittance significantly eliminates the halo artifacts in the defogged results achieved by the proposed approach,and the defogging effect is better.Finally,the proposed approach achieves excellent visual quality in comparison with other algorithms.At the same time,average gradient,information entropy and spatial frequence are employed to evaluate the defogged images quality.The experimental results show that the proposed algorithm has best evaluation results than other algorithms.