Lightweight underwater image enhancement network based on GAN

Due to the absorption and scattering of underwater light, the underwater image suffers from distortion and loss of details, which seriously affects the detection and recognition of subsequent underwater target. In this paper, a lightweight fully convolutional layer generative adversarial neural network DUnet-GAN is proposed to enhance underwater image. According to the characteristics of underwater image, this paper proposes a multi-task objective function, which enables the model to enhance the image quality by perceiving the overall content, color, local texture and style information of the image. In addition, we compare DUnet-GAN with some important existing models and make a quantitative evaluation. The results show that in EUVP dataset, the PSNR of the proposed model is above 26 dB, the SSIM is 0. 8, and the number of parameters is 11 MB, which is only 5% of the number of parameters of other models with the same performance and better than the FunIE-GAN with 26 MB parameters. Meanwhile, UIQM is 2. 85, second only to Cycle-GA N model, and the enhancement effect is significant subjectively. More importantly, the enhanced image provides better performance for underwater target detection and other models, and also meets the lightweight requirements of models for equipment such as underwater robots.