Research On Unsupervised Image Enhancement Based On Generative Adversarial Networks

Images have gradually grown in importance as a means of communication and social interaction because of the rapid development of social media and the Internet.Since lowquality photos negatively impact visual perception,more and more researchers are working to increase image quality.Image enhancement is a type of image processing that,at present,is used by social media platform software,retouching software,and built-in printer software to efficiently increase the quality of photographs and make them visually pleasing.Image enhancement algorithms have been in the limelight for a long time.Traditional techniques typically concentrate on enhancing the brightness or contrast of an image,which has a limited effect on image quality,but falls far short of satisfying human aesthetic needs.Additionally,as deep learning has advanced,many researchers have begun to concentrate on how to improve the overall visual effect of images.But they frequently rely on paired datasets for supervised learning,which makes training more challenging and has led to the gradual emergence of unsupervised image enhancement algorithms.However,such tasks encounter considerably more difficulties.On the one hand,due to the difficulty of unsupervised training,the conversion of low-quality images to high-quality images requires guidance information,which is difficult to determine.On the other hand,it is crucial to explore and make full use of the information in the image itself to achieve overall image quality improvement.To address these issues,this thesis studies and explores image enhancement tasks from two perspectives of "overall image style and color transfer" and "core image attribute enhancement".(1)Based on the idea of image style transfer,this thesis views image enhancement as a conversion from low-quality to high-quality domain.And proposes a novel bidirectional normalization and color attention-guided generative adversarial network(BNCAGAN),which provides guidance for the conversion process through high-quality images and generates enhanced images based on generative adversarial networks.Specifically,the style information of the high-quality images is first extracted and fully integrated with the content information of the low-quality images,while an inverse structure is designed to strengthen the stability of the network.In addition,the attention of the generator to the hard and difficult parts is enhanced by an auxiliary classifier,while the discriminator enhances the attention to the color channel in an attention mechanism to distinguish the low and high-quality images in terms of color style,thus encouraging the generator to produce realistic and natural visual effects.With the limitations of adversarial loss,feature-level loss,and classification loss,this structure can improve the quality.(2)This thesis suggests a novel core-attributes enhanced generative adversarial network(CAE-GAN)to achieve quality improvement by enhancing the attributes of an image one at a time from the viewpoint that various attributes of an image have a significant impact on its quality.Brightness,sharpness,and color are three attributes that this thesis specifically chooses as being crucial for photographs.To increase brightness pixel by pixel,the brightness of the image itself is first employed as an a priori to reflect the degree to which certain parts need to be brightened.Second,by adding back the highfrequency edge information lost during down-sampling in the reconstruction process,the visual detail of the image is enhanced.Finally,the discriminator enhances the control of global image tones from the perspective of statistical characteristics.This architecture significantly improves the overall image quality under the combined constraints of improved adversarial loss,feature-level loss,and pixel-level loss.In this thesis,we undertake extensive experiments on multiple publicly available large datasets to demonstrate the effectiveness of the above approach.Meanwhile,the advantages of the image enhancement algorithm proposed in this thesis is verified by comparing it with advanced algorithms in terms of objective evaluation metrics and subjective visual perception.