Single Low-light Image Enhancement Based On Multi-Scale Fusion And Deep Learning

Due to the influence of the acquisition environment,for example,under the conditions of low illumination,such as cloudy days,nights,and objects which are blocked,the images obtained by the acquisition equipment are often of low contrast,high noise,and severe loss of detail information.Some image enhancement algorithms like Multi-scale exposure fusion which cannot enhance image details accurately,nor can they recover image color precisely.In recent years,it is significantly developing in deep learning and widely used in the field of image processing.Compared with many traditional algorithms,a deep learning method can learn prior knowledge from massive training sets and has stronger representation ability.It can adapt to different scenes by adjusting parameters.Go further,the algorithm based on deep learning requires less online computing power of the hardware system with inputting the image to be processed and the corresponding real image of the ground.This thesis proposes a hybrid learning approach for single low-light image enhancement in view of the advantages and disade of the methods mentioned above.The method is composed of three steps,the first step includes the measurement of two virtual images,medium and high-exposure,by IMF as traditional color distortion tends to be induced in undereposed areas.Therefore,in order to create the virtual image,edge protecting filters(like weighted directed image filters and gradient-dominance directed image filters)are applied if the pixel value is less than the threshold value of at least one line.When the pixel value of each channel reaches the threshold value,the IMF directly generates virtual pixels.The second step is to let the loss function learn the difference value,which is generated by the difference between two images with middle and high exposure ratio and the corresponding real image,and through the neural network,the enhanced virtual image will have arrived.Finally,a resulting image is generated by multi-scale exposure fusion.In this thesis,it introduced traditional algorithms and the application of deep learning model for processing low-light images in chapter two.Then,chapter three suggests a hybrid learn-ing method based on multi-scale fusion.The method of deep learning mainly uses the loss function to refine the intermediate error,improve the two virtual images with middle and high exposure ratio,and finally produce an image with less color distortion and robust detail optimization through multi-scale fusion.Through several experimental tests,this proposal compares the results of the current ad-vanced algorithm.And the proposed algorithm will help minimize color distortion and make the image information more visible compared to the conventional image processing approach.Compared with other deep learning low-light processing algorithms,the conver-gence speed and accuracy of the proposed algorithm are significantly improved.