基于局部生成对抗网络的水上低照度图像增强

To address the problems of the maritime images taken in low-light, including low brightness, low contrast and poor quality, this paper proposes an image enhancement method based on a local Generative Adversarial Network(GAN).The generator is designed by taking the residual network as the backbone, and a pyramid dilated convolution module is introduced to extract and learn the deep features and multi-scale spatial features of images, reducing the loss of structure information.At the same time, an autoencoder is designed as an attention network to estimate the light distribution of the image and guide the adaptive enhancement for regions of different brightness.Finally, a discriminator that is able to distinguish local regions of the image is designed to constrain the generator to output images with more natural enhancement effects.Experimental results show that the proposed method can effectively enhance maritime images taken in low-light.Compared with SRIE, LIME and other traditional methods, the proposed method can restore scenes better and retain more details.