| Image-to-image translation is a field of image processing in computer vision,the purpose of which is to train a model for image-to-image translation.Image enhancement is a direction of image-to-image translation,although methods based on deep learning have achieved a lot in image enhancement and image restoration,there are still some shortcomings.This paper proposes a neural network for image enhancement based on generative adversarial network(GAN),and studies that when normal/low-light image pairs are lacking,it is still possible to learn the mapping from low-light images to normal-light images,because it is very challenging to obtain such image pairs under normal conditions.This paper conducts research on real image datasets,and mainly solves the following core problems:using GAN network to train on unpaired datasets to enhance low-light images;at present,there are many problems in low light image enhancement,such as color distortion and lack of image details;most Retinex algorithms use the reflection image as the final result,the naturalness of the generated image will not be enough,so a more suitable Retinex algorithm should be introduced into the neural network to decompose the image.In view of these three directions,this paper proposes the following innovations:(1)This paper proposes a GAN network that can be used for low-light image enhancement of unpaired images,in which the direction of image generation is fixed,the content in the generated image is consistent,and the illumination of the image is restored.(2)The generator of the GAN network is improved,and a skip connection of the attention mechanism is added to the generator to guide the generation direction of the generator,and a new loss function is proposed to train the entire network.(3)A more suitable Retinex image decomposition algorithm is proposed,which can estimate the illumination component from the image,and fully use the illumination component as the input of the network to learn the mapping. |
PDF Full Text Request