Efficient Algorithms For Image Quality Enhancement

Image is captured for the representation of visual information and used very frequently in today's life.The distortion of practical image application often induces distortion to image.To solve this problem,image quality enhancement method was proposed and used to produce the high-quality image from the low-quality one.Enhancing image quality has been the most challenge problem in image processing.However,the traditional enhancement method based on the classical signal processing theory has encountered the technique bottleneck and their performances are limited.Recently,convolutional neural network?CNN?have been successfully applied to many practical applications of image processing.In this thesis,we aim at designing the CNN-based image quality enhancement algorithms for various applications,including the proposed three-stage deep convolutional image demosaicing algorithm,the deep convolutional reconstruction for the transform domain down-sampling-based color image compression,and the bitmap-based image deblocking network.The three-stage deep convolutional image demosaicing algorithm uses CNNs to fully explore the correlation between channels,and redefines the three-stage process of image demosaicing to achieve higher demosaic performance.The proposed CNNs are mainly composed of prior feature fusion units?PFU?,adaptive local residual units?RRU?and Laplace energy constrained local residual units?CRU?.The idea of guided filter is used for the fusion of G-channel features and R/B-channel features in the PFU,it makes full use of the prior information to reduce the difficulty of network modeling.The residual mechanism with self-correction is used to reduce the error of each residual block and improve the accuracy of network modeling in the RRU.The CRU refines the learning target of the specific residual blocks in the network and enhances the main information in the residual features.The loss function combining the pixel domain and the transform domain is used in the deep convolutional reconstruction for the transform domain down-sampling-based color image compression to establish an image restoration CNN.The loss function can not only guarantee the objective quality of the reconstructed image,but also improve the fidelity of the reconstructed image details,so that the algorithm can recover the high-quality image without high frequency coefficients.In addition,the improved transform domain down-sampling algorithm based on ₁l norm and ₂l norm joint optimization and the proposed image restoration algorithm are applied to the compression of chrominance components.A new color image coding is constructed,which improves the coding efficiency of traditional algorithms.The bitmap-based image deblocking network is used to improve the quality of compressed images,mainly composed of frequency domain separated feature enhancement units?FSU?.In the FSU,the high-frequency part and low-frequency part of the features are first separated by discrete cosine transform?DCT?,then the"Squeeze-and-Excitation"mechanism is used to enhance the feature representation,and finally the feature important matrix extracted from bitmap is used to recalibrate features by attention mechanism.In this way,not only the combination of encoded information and image information is achieved,but also the specificity and interpretability of the convolutional kernels are increased.