Research On Low-Light Image Enhancement Methods Based On Generative Adversarial Network

Images play a pivotal role in people's daily life and are one of the main sources for people to obtain information.When people use cameras or phones to take photos,due to the brightness of scene and shooting equipments,these images will have problems such as low contrast,poor visibility,and high ISO noise.Such image detail information is not obvious,which reduces the practicability of them.Therefore,people have proposed a large number of methods to change the contrast of images,make their details obvious,and improve their quality and use value.When some methods enhance images,problems such as overenhancement and partial loss of details usually occur.Moreover,many of them rely heavily on paired low/normal-light images when training models,and will lose their original advantages if they do not.This topic focuses on the research of low-light images enhancement methods.It aims to improve the use value of low-light images by studying some technologies and methods,and make the details of dark area more obvious and have better visual effects.Thesis uses Generative Adversarial Network(GAN)as the main research method to propose the following two low-light image enhancement methods:(1)In order to solve the problems of excessive enhancement,partial loss of details,color and feature distortion,thesis proposes a supervised low-light image enhancement method based on GAN with three discriminators on the basis of improving an image conversion model,referred to as SFPGAN.Firstly,in order to make enhanced images retain more details and be closer to real images,color discriminator,gray discriminator and gradient discriminator are used to distinguish the trueness of generated images from three directions of color,brightness and texture.Secondly,in order to preserve the features of lowlight images,the self-feature preservation loss is introduced.Finally,in order to make the model more robust,it is trained with a set of images containing a certain amount of normal brightness and overexposure images.(2)By introducing self-correcting attention mechanism module in color generator,an unsupervised low-light image enhancement method based on the GAN attention mechanism is proposed,referred to as SAMGAN.Firstly,two different generators are used to enhance the color and grayscale images of original images.At the same time,two discriminators are used to judge their authenticity.The attention map of low-light images is used in color generator as a self-regularized attention mechanism.Secondly,in order to preserve the features and content of low-light images,the self-feature preservation loss is introduced.Finally,the gray-level consistency loss is used to ensure that the gray-level images of images generated by the two generators are consistent.After a large number of experiments,it is proved that the proposed methods can effectively solve the above-mentioned problems.In the subjective and objective evaluation of images,they are superior to some typical methods.They not noly have more pleasing visual effects,but also can enhance low-light images in real world.SFPGAN can retain the semantic information of low-light images.In the objective index of image evaluation,the performance of SFPGAN is better than that of SAMGAN by comparing the two methods.