Research On Edge-aware Image Filtering Based On Soft Clustering

Edge-aware image filtering is a significant technology in computational photography.The purpose is to perceive and retain essential edge features in the image when performing filtering,called edge-preserving filtering.According to the different calculation methods,traditional edge-aware filtering can be divided into local and global methods.The output of the local method is calculated from the weighted average of pixels in the local neighborhood,and the global method filters image by modeling and solving the optimization problems.Most of the existing local methods have high computational efficiency but insufficient edge-preserving ability,which are prone to introduce artifacts such as halos.Most of the current global methods have a robust edgepreserving filtering ability.However,because of needing to solve complex models,they are computationally expensive and even introduce artifacts such as intensity shifts.The effect and efficiency of edge-aware filtering need to be improved urgently.The edge generally refers to the area where the pixel gradient changes sharply.In the scene with rich textures,the textures will be mixed with the structural edges,and the traditional gradient-based filtering algorithm will no longer be applicable.Structure-preserving filtering is proposed to extract meaningful structural edges from fine-scale textures.Because of the need for more vital edge-preserving ability,structure-preserving filterings have low computational efficiency;most existing methods have unsatisfactory structure-preserving and texture filtering effects.This thesis proposes an edge-aware image filtering based on soft clustering,and the specific work is as follows:(1)For suppressing artifacts and improving edge-preserving filtering ability and computational efficiency,an edge-preserving filtering based on soft clustering is proposed,combining edge-aware tasks with soft clustering.A soft clustering algorithm based on a limited form Gaussian mixture model is proposed to avoid the defects caused by embedding existing methods.According to the given parameters,soft clustering pixels can obtain the partition matrix,and then the affinity matrix can be constructed to perform filtering.The method can suppress intensity shift because the output is a weighted average of pixels in a local window;the method can suppress halo because the weights derived from soft clustering separate different pixels.The amount of smoothing is lacking in deep learning-based methods,which can be solved by providing flexible parameter control in the proposed method.In terms of efficiency,the method has linear time complexity.Experimental results show that our method is beneficial for various computational photography applications.(2)For improving the structure-preserving filtering ability,a structure-preserving filter based on power iteration soft clustering is proposed by modeling the task of structure-preserving filtering as a power iteration soft clustering problem.The kernel is related to the diffusion distance,making the proposed method easy to distinguish structure and texture.The algorithm consists of two iterative processes: the former smooths out the texture by power iteration filtering the input;the latter performs joint soft clustering filtering to recover the underlying structure and shadows.The method is computationally efficient,requiring only linear time per iteration.Both qualitative and quantitative experiments demonstrate the advanced character of the proposed method.(3)Based on the above research,the edge-aware filtering image processing system based on soft clustering is designed and implemented.The prototype system embedding the above two methods displays various computational photography applications in real-time according to specified parameters.The system comprehensively covers the basic functions,proving the theoretical significance and practical value of the research in this thesis.