Research On Two-Stage Low-Light Image Enhancement Based On Pre-Enhancement And Post-Restoration

In low-light scenes such as night,only a small amount of light enters the camera,resulting in under-exposure,which greatly affects the visual effect of images.In order to solve the above problems,treatment is needed to enhance the low-light images,so that the enhanced images have similar subjective experience and objective indicators of clear images with normal light.Existing methods for enhancing low-light images suffer from noise and color distortion.In order to improve the quality of enhanced images,this article proposes a low-light image enhancement method based on enhancement followed by restoration.This method first enhances the illumination of the low-light image and then restores the noise and color distortion in the enhanced image.The main research content and innovation of this article are as follows:First of all,in response to the problem of weak brightness of the low-light images,this article proposes an image enhancement algorithm(Singular Value Enhancement,SVE)based on SVD(Singular Value Decomposition)transformation.SVE transforms low-light images from the spatial domain to the SVD domain,and then cuts the potential noise in Channel V by truncating the singular values.Due to the excessive compressing the dynamic range of the gray value in singular value normalization,an adaptive parameter is introduced to restrict the normalization.Finally,transform the image back to the spatial domain and use the "Contrast Limited Adaptive Histogram Equalization" to enhance the image contrast.The experimental results show that compared with the original singular value normalization method,the SVE algorithm can effectively reveal the details of the dark regions and make the enhanced images have higher brightness and contrast.Then,in response to the problem of weak brightness of the low-light images and the low speed of SVE algorithm,this article proposes an image enhancement algorithm(DWT-QR)based on DWT-QR transformation.DWT-QR uses DWT(Discrete Wavelet Transform)to obtain the LL sub-band representing the light information in the low-light images,and then the QR decomposition is performed on the LL sub-band to get the upper triangle matrix to adjust the light intensity.The brightness reference image used to obtain the upper triangular matrix enhancement coefficient is obtained by implementing adaptive gamma correction through the LL sub-band.Because the size of the LL sub-band is only a quarter of the original picture,and the QR decomposition is faster than the SVD,the computation complexity of the DWT-QR is much smaller than SVE.The experimental results show that the DWT-QR is significantly faster than SVE,and it can retain the details well.Finally,in response to the problem of noise and color distortion in the SVE and DWT-QR results,this article proposes the RestoreNet based on SE-Block(Squeeze and Excitation Block)and U-net.RestoreNet consists of two parts,which are SE-Block&U-net based generator and discriminator.The U-net extracts the multi-layered features of the images by encoding-decoding structure,and reconstructs the images of the same size with higher quality.The SE-Block is added between the skip connection of the U-net,so that the adverse features produced during the convolutional process are filtered as much as possible,and the channels that make the output images of high quality are valued.The discriminator will judge the output of the generator and the real high-quality images,and use the judgement as a loss function to guide the generator to produce images with higher quality.The experimental results show that RestoreNet can better reduce noise and color distortion in SVE and DWT-QR results compared to other image restoration measures.