基于硬负样本对比学习的水下图像生成方法  

Underwater Image Generation Method Based on Contrastive Learning with Hard Negative Samples

作　　者：刘子健 王兴梅[1,2] 陈伟京 张万松 张天姿[1] LIU Zijian;WANG Xingmei;CHEN Weijing;ZHANG Wansong;ZHANG Tianzi(College of Computer Science and Technology,Harbin Engineering University,Harbin 150001;National Key Laboratory of Underwater Acoustic Technology,Harbin Engineering University,Harbin 150001)

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,哈尔滨150001 [2]哈尔滨工程大学水声技术全国重点实验室,哈尔滨150001

出　　处：《模式识别与人工智能》2024年第10期887-909,共23页Pattern Recognition and Artificial Intelligence

基　　金：水声技术重点实验室稳定支持课题(No.JCKYS2024604SS JS006);中央高校基本科研业务费专项资金(No.3072024XX 0602)资助。

摘　　要：在获取稀缺水下图像时图像生成技术至关重要,通常依赖有序配对数据.考虑到实际海洋环境中获取该类数据受限,引入CL-GAN(Contrastive Learning-Based Generative Adversarial Network),克服图像域双射条件的限制,但由于随机采样的负样本质量较低,模型难以从水下噪声图像中学习复杂内容特征.因此,文中提出基于硬负样本对比学习的特征级生成对抗网络(Hard Negative Sample Contrastive Learning-Based Feature Level Generative Adversa-rial Network,HCFGAN),用于水下图像生成.为了提高负样本质量,提出硬负样本采样模块(Hard Negative Sampling Module,HNS),挖掘样本间的特征相似性,将靠近锚点样本的硬负样本加入对比损失中,学习复杂特征.为了保证负样本的复杂性和全面性,构造负样本生成模块(Negative Sample Generation Module,NSG).通过NSG和HNS的对抗性训练,确保硬负样本的有效性.为了提高模型对水下模糊图像的特征提取能力及训练稳定性,设计上下文特征生成器和全局特征判别器,增强对细微内容特征和水下风格信息的感知能力.实验表明,HCFGAN生成的水下图像具有良好的真实性和丰富性,在水下图像生成实际应用中具有重要价值.Image generation is essential to acquire scarce underwater images,and it is typically reliant on paired data.Considering the limitation of practical access to such data distributions in marine environments,a contrastive learning-based generative adversarial network(CL-GAN)is introduced to overcome the constraints of bijection in image domain.However,the model struggles to learn complex content features from noisy images due to the low quality of negative samples resulting from random sampling.To address this issue,a hard negative sample contrastive learning-based feature level GAN(HCFGAN)for underwater image generation is proposed.To improve the quality of negative samples,a hard negative sampling module(HNS)is designed to mine feature similarity between samples.The hard negative samples close to the anchor sample are incorporated into contrastive loss for complex feature learning.To ensure the complexity and comprehensiveness of negative samples,a negative sample generation module(NSG)is constructed.The adversarial training of NSG and HNS ensures the validity of hard negative samples.To enhance feature extraction capability and training stability of the model for underwater fuzzy images,a contextual feature generator and a global feature discriminator are designed.Experiments show that the underwater images generated by HCFGAN exhibit good authenticity and richness with practical value in underwater image generation.

关 键 词：水下图像生成 生成对抗网络(GAN) 对比学习 硬负样本 特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]