Research Of Deep Learning-based Low-light Image Enhancement Algorithm

Under poor lighting conditions,due to insufficient exposure,the images collected by optical imaging equipment have dim vision,blurred details and poor visibility,so they must be enhanced before they can be used for subsequent image processing tasks.Existing algorithms of low-light image enhancement(LLIE)are mostly based on Retinex optical physical model,and the images obtained have good visual effect on the whole.However,it is easy to cause local detail blur problem when estimating irradiation or reflection components.Recently,LLIE algorithm based on depth learning has become a hot topic for researchers.This kind of low-light image enhancement algorithm has some advantages in image contrast and detail enhancement for specific image content.However,due to the inherent data dependence,not all given low illumination images have a stable enhancement effect.Based on the in-depth analysis of the advantages and disadvantages of various implementation strategies of the mainstream low-light image enhancement algorithm,and inspired by the fusion framework,this paper proposed a LLIE algorithm based on multi-image local structured fusion-based(MLSF)LLIE.In the pseudo-exposure image preparation stage,we first utilized the pre-trained prediction model to estimate the optimal exposure ratio of a given low-light image.Based on this,we can exploit the brightness transform function(BTF)to generate a well-exposed image by feeding the BTF function with the estimated optimal exposure ratio.Moreover,in the same manner,we also generate an over-exposed image with.Secondly,in order to better the local structure details in the image texture and color fidelity for preserving effect,using classical Retinex model to generate a moderately enhance image as a supplement images involved in fusion,the low-light image,two well-exposed images and the over-exposed image at the same spatial positions were vectorized and decomposed into independent components,i.e.,contrast,texture structure,and brightness.For the contrast component,the maximum contrast was used as the contrast component of the fused patch,while for the brightness and structural strength information the phase consistency and the visual saliency were used as the weight coefficient respectively,during the fusion.Finally,the enhanced image was obtained by reconstruction of patches based on the three fused components.In view of the powerful and automatic feature extraction function of DCNN,it can make full use of the prior knowledge contained in the natural image to improve and enhance the effect.Based on deep learning and image fusion(DLIF),this cites proposes a low illumination image enhancement algorithm based on hybrid strategy.We first adopted deep learning to train an illumination prediction model,which can quickly estimate the optimal illumination component from a given low-light image and obtain its corresponding moderately exposed image in the framework of the Retinex model.Considering the fact that the under-exposed and over-exposed areas may still exist in the moderately exposed image,the low-light image and its over-exposed image were used as supplementary images for the moderately exposed image.Finally,the three images to be fused were converted from RGB to YCb Cr and the Y component of the low,moderate,over-exposed images were fused within the framework of local structured fusion.At the same time,the chrominance weighted fusion mechanism was used to fuse the Cb and Cr components of the low,moderate,and over-exposed images,respectively,then the fused Y,Cb and Cr channels were combined and converted back into the RGB color space to obtain the final enhanced image.In order to verify the proposed MLSF and DLIF algorithms,comparative experiments with the classic LLIE algorithm are carried out on the reference image set and the commonly used non-reference image set respectively.The two evaluation indicators of IW-PSNR(information content weighted peak signal to noise ratio)and IW-SSIM(information content weighted SSIM)are selected for comparison with reference image sets.For non-reference image sets,contrast distortion metric(noreference image quality metric for contrast distortion(NIQMC),blind image quality measure of enhanced(BIQME),optimized BIQME-optimized image enhancement method(BOIEM)and integrated local natural images A total of 4 evaluation indexes of integrated local natural image quality evaluator(ILNIQE)are compared.Among them,the larger the index value of IW-PSNR,IW-SSIM,BOIEM,BIQME and NIQMC,the better the image quality,while the smaller the ILNIQE,the better the image quality.Experimental data shows that compared with various existing mainstream LLIE algorithms,MLSF and DLIF algorithms are ranked in the top two in various evaluation indicators,which can effectively improve the visual effect of low-illuminance images obtained indoor and outdoor,and avoid Obvious color distortion,overexposure,etc.,have better image detail retention ability and color fidelity effect in the details of local image structure.