基于GAN的轻量级水下图像增强网络

Lightweight underwater image enhancement network based on GAN

由于光在水下存在吸收与散射,导致水下图像存在颜色失真和细节损失,严重影响了后续水下目标的检测和识别。本文提出了一种轻量级全卷积层的生成式对抗神经网络模型(DUnet-GAN)来增强水下图像。针对水下图像的特点,提出了多任务目标函数,使得模型从感知图像的整体内容、颜色、局部纹理和风格信息等方面来增强图像的质量。此外,与现有的一些重要的模型做了对比,进行了定量的评估。结果表明,在EUVP数据集中本文所提模型峰值信噪比在26 dB以上,结构相似度为0.8,参数量为11 MB,仅为其他达到同等性能模型参数量的5%且比26 MB参数量的FUNIE-GAN指标更好。同时UIQM为2.85,仅次于Cycle-GAN模型,且主观增强效果显著。更重要的是,增强后的图像为水下目标检测等模型提供了更好的性能,也满足了水下机器人等设备对模型的轻量化要求。

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-GAN 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.