Research On Image Enhancement Via Layer Decomposition

As cameras,smart phones,and other mobile devices spread,people can take pictures and videos anytime and anywhere.However,in the process of image acquisition,transmission and storage,there are always unavoidable disturbances such as jitter,exposure,bad weather,which degrade the quality of the images.Therefore,the enhancement of these low-quality images will help people better understand their contents.In addition,image enhancement can be used as a pre-processing for image segmentation,image recognition,texture replacement and similarity evaluation between virtual and real scenes.At the same time,the natural image enhancement algorithms need more suitable evaluation metrics.It not only needs to improve the sharpness and visual authenticity of the enhanced image,but also needs to ensure the image integrity.This thesis intends to study the enhancement of low-quality images,using different knowledge of different image layers,for restoring the content of low-quality images.We take the deraining application as an example,a comprehensive evaluation of image enhancement on deraining algorithms is carried out,and we build a bridge between low-level image enhancement algorithms and higher-level computer vision algorithms.The main contributions of this thesis can be summarized as follows:1)An image enhancement method based on structure layer prior is proposed.We have statistics the distribution of gradient field of natural scene images.We find that natural scene images has a gradient sparse feature.Then we use the gradient inconsistency and gradient independent between noise and background layers as our prior,and use the Augmented Lagrange Multiplier to solve this problem.The proposed image enhancement algorithm can be applied to image block effect and introduced by DCT based compression algorithms and rainy image enhancement.2)We design a decomposition-composition deep convolution network for image enhancement,which can separate the noise from low-quality images.When we take pictures in rainy days,there are rain streaks and haze existing in the images.We use the proposed image decomposition-composition network to decompose the rainy image into clear background and rain layer,and retain the original structure of the image after rain removal.At the same time,this method can be applied to other two-layer decomposition image enhancement problem,such as image defogging and dust removal.Considering the fact that there are fewer open source rainy data sets,we propose a new dataset,and compare our algorithm with the current rain removal algorithms on this dataset.3)In order to evaluate the image enhancement algorithms,researchers often adopt full-reference evaluation methods.These methods can not fully reflect the effectiveness of the enhancement algorithm.In view of this,we propose a new rain dataset and labeled the objects in the dataset.We also evaluate current deraining algorithms in terms of their impact on subsequent object detection tasks,as a “task-specific” evaluation criterion.We reveal the performance gap in various aspects,when these algorithms are applied on synthetic and real images.By extensively comparing the state-of-the-art single image deraining algorithms on the this dataset,we gain insights into new research directions for image enhancement.