L₀ Optimization And Its Application In Image Smoothing And Image Deblurring

With the rapid development of computer science and technology,images have become an effective way for people to perceive and understand the world.However,since the image is susceptible to external interference and errors during the acquisition and transmission process,the subsequent processing of the image becomes very difficult.Therefore,the research on image preprocessing(including image smoothing,image denoising,image deblurring,etc.)is particularly important.With the development of image processing,achieving feature selection through sparse representation of images has become more and more common.Especially after the L₀norm is used for sparse expression,many scholars have applied it to suitable fields.Among them,the L₀gradient minimization(LGM)proposed by Xu et al.in 2011 successfully applied it to image smoothing.This paper takes advantage of the L₀norm in extracting and maintaining image features,improves LGM,and applies it to the fields of image smoothing,image denoising,and image deblurring.First of all,this paper proposes a L₀ optimization algorithm based on Laplacian operator.Compared with the problem of uneven image color transition in the previous algorithm,a more effective algorithm is proposed and applied to image smoothing and image denoising.This method uses the Laplacian operator to constrain the color change of the image,and optimizes the L₀model to slow down the change of the color gradient to achieve the smooth transition of the image color.At the same time,the Sobel operator is introduced to better maintain image edge features during the smoothing process.The experimental results show that this method can reduce the loss of image detail features while smoothing the image,effectively deal with the phenomenon of stepped edges and color block distribution in image smoothing,and can effectively remove a variety of noises in the image.The peak signal-to-noise ratio and running time of the method are also improved.Secondly,a multi-scale recurrent network based on sparse expression of L₀is proposed and applied to image deblurring.The blind deblurring method based on machine learning can effectively process blurred images in the real world.However,the existing multi-level architecture may cause problems such as the inability to preserve edges,and the expected introduction of artifacts and ghosts when removing blur.This paper proposes an edge extraction module,which is embedded in a multi-scale recurrent network.It can preserve the edges of blurred images.Specifically,SRN has three levels of subnets,and each level uses an encoder-decoder structure.When the current scale transmits information to the next scale,the edge extraction module is used to perform edge enhancement.Then pass it down to the next scale.In addition,we introduce the dual attention mechanism of space and channel into the encoder-decoder structure,taking into account the correlation between pixels and the correlation between channels,and use it as new training information.The experimental results on the GOPRO dataset show that our method produces better results both quantitatively and qualitatively compared with other state-of-the-art methods.