Research And Implementation Of Dark Image Denoising And Enhancement Algorithm Based On Generative Adversarial Networks

Image denoising and enhancement have always been the key technologies to improve image quality in the field of image processing.Today,the application field of image processing is expanding,and the requirements for image quality are also improving.But in the process of image acquisition,The stability and high quality of the collected image quality can hardly be guaranteed due to the influence of equipment,environment,illumination and other factors.Therefore,image denoising and enhancement can effectively improve the image quality,which is of great significance for better image processing in the later stage.Chen et al.proposed an algorithm to denoise and enhance the low-light images collected under very weak light conditions,which can make the processed images present good visual perception.However,their algorithms directly use U-Net to generate images,without considering that different acceptance domains of different scales may lead to inconsistency of extracted features and the feature information of some detail parts is not extracted completely,which results in unclear details and discontinuous chromatic aberration blocks in the generated image.Therefore,this thesis explores and experiments the denoising and enhancement algorithm of low-light image.The main work includes the following aspects:1.Using the Generative Adversarial Network as the whole network structure,U-Net as the basic framework of the generator,MAE and the adversarial loss as the total loss of the network,to ensure that the image generated by the network is more real,effectively retain the texture boundary and tone brightness of the feature informat:ion in the real image.2,The thesis puts forward Deep Residual Inception Module and Channel Attention Block.To increase the Deep Residual Inception Module(DRI)in the contraction path of U-Net,so that the network can extract the content information of the image from more scales,and ensure the feature integrity of the generated image;In the extended path of U-Net,Channel Attention Block(CAB)is added.According to the high-level semantic information,weight vectors of different features in the shallow information are obtained,so as to guide the fusion of global and local image information and ensure the semantic consistency of generated images.It can be seen from the experimental results that the algorithm proposed in this paper can effectively remove the real noise of dark images,enhance the brightness and contrast brightness,which is significantly improved compared with Chen et.al algorithm.Besides improving the clarity of images,the algorithm can effectively restore details and solve the problem of discontinuous color difference blocks.