Research On Low-Light Image Enhancement And Image Fusion Based On Non-Local Method

As the basic research in the field of image processing research,low-light image enhancement and image fusion have been widely used in computer vision,pattern recognition,artificial intelligence and real life.In recent years,non-local ideas have achieved success in real-world image denoising.Among them,BM3 D and NLH algorithms provide the feasibility of block-level non-local and pixel-level non-local ideas in image denoising tasks,respectively.However,current non-local methods have not been effectively applied in the field of image enhancement and image fusion.This paper mainly makes the following two works:Firstly,this paper applies the block-level non-local idea in block-matching 3D transform(BM3D)to low-light image enhancement research,and proposes a Retinex model with 3D Double Transformed Decomposition.In this method,the low-light image is decomposed by 3D double transform based on block matching,and the decomposed result is synthesized by the Retinex model to obtain the enhanced image.In the three channels of RGB space,the low-light image is decomposed by 3D double transform based on block matching to obtain three groups of illumination components and reflection components.The illumination component is reconstructed from the low-frequency coefficients after 3D double transform based on biorthogonal wavelet and Haar transform.The reflection component is reconstructed from the high-frequency coefficients of the 3D double transform based on discrete cosine and Haar transform.Among the three groups of illumination and reflection components,a pair of illumination and reflection components is selected according to the minimum reflectionmaximum illumination selection strategy.Exponential and logarithmic hybrid transformation is used to enhance the illumination component,and the dot product of illumination and reflection is used to replace the V channel component in the HSV space of the low-light image.Finally,the enhanced result image is obtained by converting from the HSV space to the RGB space.Secondly,this paper applies the idea of pixel-level non-local Haar transform(NLH)to image fusion research,and proposes a Pixel-level Non-local Space-frequency Combination Based Universal Image Fusion.Block matching and row matching are performed in the image A,and block extraction and line extraction are performed in the image B,to obtain two groups of similar pixel groups.The spatial domain fusion area and the frequency domain fusion area of the image are divided by the standard deviation ratio of the two groups of similar pixels and the threshold parameter.Pixel fusion is adopted in the fusion area of the spatial domain,that is,a similar pixel group with a larger standard deviation is directly selected as a similar pixel group after fusion.In the fusion area of the frequency domain,coefficient fusion is adopted,and two sets of transformation coefficient matrices are obtained by performing separable Haar transformation on two groups of similar pixel matrices.The average coefficient value of the group of low-frequency coefficients is used as the low-frequency coefficients after fusion,and the similar pixel groups after fusion are reconstructed by combining the high-frequency coefficients after fusion and the low-frequency coefficients after fusion.Finally,the final fused image is generated by aggregating all fused groups of similar pixels.This paper further introduces a pixel-level non-local distance map to detect the edge information of the image,adjusts the size of the edge area block size according to the energy of the distance map,uses small-sized blocks in the image edge area,and uses large-sized blocks in the smooth area.So as to achieve the purpose of alleviating the block artifact in the edge area of the fused image.In this thesis,a lot of image enhancement and image fusion experiments have been carried out on several datasets.The experimental results show that the two methods proposed in this paper are better to the other methods in both subjective vision and objective evaluation.