Image Enhancement And Reconstruction Based On Modeling And Deep Learning

Image enhancement and reconstruction is a kind of basic research problem in image pro-cessing,which has been receiving more and more attention.Imaging process by equipment and scene conditions are important reasons being responsible for poor image quality,such as system noise,camera shakes,bad weather.The degraded images appeared as noisy images,blurred images,rainy or hazy images,etc.It is challenging to reconstruct visually clear images from the observed degraded images.Image enhancement and restoration is usually the fundamental low-level processing step in many practical vision systems,which aims to improve the visual quality and provide reliable in-formation for subsequent visual decisions.However,most conventional image enhancement and restoration methods focus on suppressing artifacts,fixing the relative scale ambiguity between unknowns in the degradation model,etc,there are still some disadvantages.And newly emerg-ing deep learning based methods usually pay attention to designing end-to-end task-specific network structures,which can avoid the formulation process and hand-crafting priors.Never-theless,they usually require training on the large datasets,its generalization ability is poor and the computing cost is expensive.Hence,proposing effective algorithms is an important research topic in computer vision and image processing communities.This thesis mainly focuses on four image enhancement and restoration problems,including edge-preserving filtering enhancement,single image deblurring,single image dehazing,and single image deraining.Four algorithms base on the model and deep learning are proposed to handle with the four corresponding problems.The main contents of this thesis are as follows:(1)Image enhancement and reconstruction based on adaptive regularization model.We propose a edge-preserving filtering method for image enhancement,which helps effective-ly avoid the structure loss and staircase effect.To address adopting hard threshold by the existed L0 gradient minimization method,we propose to formulate an adaptive L0 gradient minimization model based on the feature-driven.The proposed adaptive function makes the utmost of the image gradient,improving the ability of the filter to remove details and preserve key structures.The proposed method is able to generate edge-preserving output-s and overcome staircase effect,which can avoid introducing unpleasing effects in some enhancement applications.In addition,we analyze the relative scale ambiguity between the latent image and the blur kernel and propose a scale normalization method for single blind deblurring based on the Lp regularization constraint on the blur kernel.We show that the hyper-Laplacian regularizer can be transformed into a joint regularized prior based on a scale factor.And the weight coefficient of the regular term is no longer a fixed scalar,it changes adaptively in the optimization process.The proposed model can be adopted to both kernel estimation and intermediate latent image reconstruction.Experimental re-sults demonstrate the proposed methods can be applied to many image enhancement and reconstruction problems.And it can achieve good performance.(2)Single image dehazing based on weakly supervised network.We propose a weakly su-pervised network for single image dehazing.The proposed network only uses the ground truth of haze-free image for supervision and automatically estimates the transmission map and the atmospheric light.It generates reliable restoration of haze-free images by em-bedding the physical-model.In addition,utilizing the real-world dataset can fine-tune the network,which improves the performance on the real-world images.Experimental results demonstrate the proposed method achieves good performance in single image dehazing.(3)Single image deraining based on densely connected pyramid network.We propose a densely connected pyramid network for single image deraining.Based on the analysis that multi-scale information is useful for extracting rain components,the pyramid polling block is embed following the dense connection.Employing the dense connection can maximize the information flow along features from different levels,which enhance the learning ability of the features.The multi-scale pyramid pooling module is adopted for improving the capture of smulti-scale information,which can further facilitate to extract rain components.Experimental results demonstrate the proposed method achieves good performance in single image deraining.