一种应用于Pix2Pix的水下鱼类图像增强训练集构建新方法  

A New Method for Training Set Construction of Underwater Fish Image Enhancement Applied to Pix2Pix

作　　者：李沅衡 林茂兹 丁可武 陈文辉 LI Yuanheng;LIN Maozi;DING Kewu;CHEN Wenhui(Guangdong Provincial Laboratory of Southern Marine Science and Engineering,Guangzhou,Guangdong 511458,China;Fujian Polytechnic Normal University,School of Oceanography,Fujian 350300,China;Fujian Polytechnic Normal University,Key Laboratory of Environmental Measurement,Control and Management in Coastal Waters for Fujian Institutions of Higher Learning,Fujian 350300,China;Fujian Polytechnic Normal University,Institute of Oceanography,Fuqing,Fujian 350300,China)

机构地区：[1]南方海洋科学与工程广东省实验室(广州),广东广州511458 [2]福建技术师范学院海洋学院,福建福清350300 [3]福建技术师范学院近海流域环境测控治理福建省高校重点实验室,福建福清350300 [4]福建技术师范学院海洋研究院,福建福清350300

出　　处：《福建技术师范学院学报》2024年第2期10-16,共7页JOURNAL OF FUJIAN POLYTECHNIC NORMAL UNIVERSITY

基　　金：广州市科技局基金(202201011192、202206050002);福建省自然科学基金项目(2020J05247).

摘　　要：提出一种量化对比不同参数下带色彩恢复的多尺度网膜增强(MSRCR)算法的图像增强效果,选取最优组成训练集的方法,并基于该训练集训练Pix2Pix模型.实验结果表明,利用该方法所建立的训练集,可以获得用于水下鱼类图像增强的优质Pix2Pix模型参数,且此方法还可以扩展应用于其他对抗神经网络(GAN)模型训练中.This paper proposes a method to compare the image enhancement quantization effect of multi scale retina enhancement algorithm with color restoration(MSRCR)method under different parameters by using UIQM(underater image quality measurement)and selects the best one to form a training set,which is employed to train the Pix2Pix model.Experimental results show that high-quality Pix2Pix model parameters for underwater fish image enhancement can be obtained by the training set established through this method,and the method can also be extended to other generative adversarial network(GAN)model training.

关 键 词：水下鱼类图像增强 训练集 MSRCR Pix2Pix 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]