Data Driven Low-light Image Enhancement

The images collected under low illumination environment have some degradations such as low brightness,weak contrast,noise and color bias.These degradations hinder people’s visual experience and subsequent computer vision processing and also have a negative impact on many applications such as night security monitoring and military reconnaissance.Low-light image enhancement technology can solve these degradation phenomena to some extent and help people and computers better understand image content.The low-light image enhancement algorithms based on deep learning are strongly influenced by training data.However,existing low-light image enhancement datasets only have limited annotated samples and types.In some case,the annotated samples are missing.These problems restrict the enhancement performance and generalization ability of the algorithms.In view of this problem,this thesis constructs three robust low light image enhancement algorithms by using synthesized data,paired data with different exposure levels,and unpaired data,respectively.Specific research contents and contributions are as follows:1)Low light image enhancement method based on transfer learning and adaptive group normalization.First of all,for the problem of limited annotated samples and types,the first solution proposed in this thesis is to use synthetic data to expand the number and type of dataset and improve the generalization ability of the model.For the distribution difference problem between synthetic data and real data,an adaptive group normalization method is proposed to eliminate overfitting and distribution difference among different data,and a transfer learning method is used to transfer the knowledge learned from synthetic data to real low-light data.In order to improve the enhancement performance of the algorithm,a multi-level codec network is constructed.By integrating the shallow and deep features of the network and introducing the non-local attention module to learn the global context information,the network can obtain better enhancement results.2)Light-tunable low-light image enhancement network based on Retinex model and structural prior constraints.In order to solve the groundtruth missing problem of reflection and illumination in Retinex model,this thesis proposes a layer decomposition network based on a variety of structural priori constraints to decompose the image into reflection map and illumination map.In this network,only pairs of images with different illumination levels are used as training data,which eliminates the need for groundtruth data.To solve the problem of noise and color bias in low-light images,a reflectance restoration network based on attention mechanism is proposed to suppress noise and correct color.In order to generate a variety of enhanced results,an illumination adjustment network is proposed.By learning the mapping relationship between different lights,the network can generate enhanced results with any brightness level according to user input.3)Low-light enhancement and denoising network based on unpaired data.In order to solve the problem of limited annotated samples,this thesis further proposes a low-light enhancement and denoising network based on unpaired data,which can get rid of the dependence on annotated data.In order to solve the problem of lack of constraints on unpaired data,this thesis proposes to use perceptual loss and generative adversarial loss to constrain the generated results of the algorithm.The luminance enhancement network based on the feature attention residual module and generative adversarial network can improve the brightness and contrast of low-light images.Secondly,to solve the problem of poor noise suppression performance of the brightness enhancement network,a noise block extraction algorithm is proposed and a paired noise/clear image dataset is synthesized.An image denoising network is constructed to suppress the noise of the enhanced images.In conclusion,this thesis proposes three different data-driven low-light image enhancement methods to solve the dependency of deep learning algorithms on annotated data.In addition,the proposed methods can be extended to other image enhancement and restoration tasks,such as image deraining,image dehazing and image demoiréing.