Research On Multi-functional Restoration Models Of Image Enhancement

Digital images are widely used in all aspects of daily life.High-quality digital images with detailed features such as edges,textures and high-fidelity colour provide a pleasing impression.However,capturing digital images is inevitably affected by unfavourable factors such as camera shake,uneven lighting,and bad weather etc.As a result,the acquired images are blurred,the target structure is difficult to recognize,and the visual effect is unsatisfactory.Besides,the effectiveness of computer vision applications,such as intelligent vehicles,robotics,surveillance system,object recognition,tracking,and detection etc.,are also degraded.To cope with these difficulties,image restoration has drawn a considerable amount of research interests from the computer vision community.In particular,low-light image enhancement has received substantial attention,which is attributed to its capability of revealing veiled details and tackle other associated visual aesthetic problems.Although numerous methods have been implemented for image enhancement,but these methods are limited to a single degradation type,i.e.,only enhance a particular type of images or application dependent.Also,these methods have not considered important visual appearance factors,such as simultaneously fixing high-intensity and low-intensity areas,noise reduction,colour correction,detail enhancement and reducing hazy effect.Hence,image enhancement is a highly complex process and has always been an inspiring and challenging research area.In light of the issues mentioned above,this thesis provides four innovative solutions,which handle multiple visual appearance factors while increasing the visibility of various illumination degraded images.In the below subsequent paragraph,the abstract information of each model is illustrated.(1)The first method of the dissertation investigates the enhancement of low-light image in a complex light environment.Typically,captured images in complex light conditions are obtained with low visibility,latent colour and poor contrast.Consequently,it is essential to reveal such visual aspects in order to achieve adequate visual quality and preserve naturalness.Hence,a novel model is proposed,in which it can improve the brightness,enhance contrast,process the colours,and eliminate the hazy effect up to a great extent.Accordingly,the presented model uses three main steps and other associated image processing operations to attain enhanced results.In the first step,a complete optimization function is developed to estimate the illumination component using Retinex model.Next,two functions,i.e.,camera response and bright transform,are implemented to adjust the initial enhanced results.In the second step,a nonlinear stretching function(NSF)is introduced,which plays an incredible role in our model.In NSF,the parameter ? controls brightness and contrast.For instance,lower the value of ? yields maximum stretching and higher the value of ? eliminates fog in images up to a great extent.In the third step,a cumulative distribution function of hyperbolic secant distribution(CDF-HSD)is adopted to enhance the overall visibility of dark details.The HSD is an eminent statistics and probability function.However,CDF is a class of S-curve functions,which can reshape contrast and brightness.Nevertheless,the pixel distribution of the obtained image in this step is limited to a particular dynamic range,and its visual appearance is very white.Finally,a normalization function is employed that linearly boost the pixels values.The empirical evaluation and comparison of the most recent state-of-the-art approaches on nine datasets revealed that the proposed method efficiently adjust low-light images,addressed the naturalness preservation,noise reduction and expunged hazy effect in degraded images.(2)The second method of the dissertation is developed to enhance underexposed/overexposed LDR and HDR images,respectively.In addition,it avoids unnecessary distortions in lowand high-intensity areas of degraded images.To address this issue,inspired by wavelet theory and its robust performance in translation sensitivity,the presented method for image enhancement is based on dual-tree complex wavelet transform(DT-CWT).The DT-CWT decomposes images into two sub-bands,namely low and high-frequency bands.Two algorithms are developed and applied to high and low-frequency sub-bands.An algorithm based on a fractional-order calculus is implemented to eliminate noise from the high-frequency component.It is pertinent for this model due to its vigorous performance and curbs noise up to a considerable extent.However,to boost the dim contents in degraded images,the multi-scale decomposition algorithm is chosen for the low-frequency component.Therefore,any transformation into low sub-bands will not affect the information in high-frequency sub-bands and vice versa.Moreover,this model aims to adjust the light of high-intensity regions by using sigmoid functions.Besides,the available colour cast has been discarded via white balance strategy.Moreover,the structural details are boosted via nonlinear stretching and gamma functions.The findings of extensive experiments on ten datasets showed that the proposed method accurately retains the edges,amplifies the contrast,enhances colour,less time consuming,correct high-intensity contents,and essentially eradicates noise for a visually pleasing and natural look in the images.(3)The third algorithm of the dissertation introduces a method to improve images with uneven Illumination.To enhance various illumination degraded images,the prior models are either enhance single underexposed or overexposed images or limited to a specific type of image.To fix this problem,we have drawn up a novel idea,which blends the image enhancement algorithms with machine intelligence.Firstly,Goog Le Net has been utilized to classify all kinds of degraded images into different classes such as nonuniform,side illumination,backlit,nighttime,overexposed,and slightly contrast degraded images.Next,image degradation model(IDM)is utilized to improve the visual quality of degraded images.Similarly,as the transmission map and environment light estimation are critical factors for IDM that are measured using bright channel prior and effective filter.Besides,the estimated transmission map is further refined via `1-norm regularization,which plays a vital role in producing better-perceived quality in the images.Additionally,one of the critical observations of partially underexposed images is that they have uniform brightness and contrast.Consequently,the estimated refined transmission map has been used for partially distorted low-light images in the post-processing step,from which discernible,hazy free,and detailed enhancement results are obtained.Meanwhile,colour distortion is minimized via a denoising model,specifically when increasing visibility in the extreme dark images,e.g.,nighttime or low-light images.Moreover,over-enhanced images are properly corrected via a tone mapping scheme.Experimental outcomes on nine datasets and user study showed that the enhanced results are remarkable in term of efficiency and effective for image restoration.(4)The fourth method of the dissertation is presented to improve low-light images and eliminate the hazy effect in fogy images.Most pre-existing deep learning-based image enhancement models are incapable of flexibly adjusting illumination in low-light and eliminate their hazy effect.Besides,the convolutional neural network(basic network)in prior learning-based approaches enhanced low-light images up to a constant level of brightness.To solve this issue,a novel image enhancer based on decoupling learning and the Taylor expansion(DLT-Net)is proposed.It is capable of adjusting low-light images and eliminating noise and hazy effect in the images.Accordingly,the Retinex model is modified to construct a basic network employing the attention mechanism and the Taylor expansion.The attention module eradicates undesired chromatic aberration and Taylor expansion integrates the features from various convolutional layers to predict the enhanced image.Subsequently,the weights and biases of the basic network are decoupled using a weight-bias learning network,allowing for adjustable brightness control.Moreover,a simulation method is implemented that takes into account the in-camera signal processing pipeline and the Gaussian-Poisson mixed noise model for synthesizing training data to eradicate the noise.Finally,a method of simulating realistic low-light images and the local mean and variance loss is applied to further boost the performance of the proposed DLT-Net.Extensive experiments on twelve different datasets demonstrate the qualitative and quantitative advantages of our method over cutting-edge algorithms.In addition,the potential advantages of our method for dealing with haze removal,face detection,and exposure correction are discussed.