摘要
深度卷积神经网络是水下图像增强的主要方法之一,但其过高的内存消耗和计算需求阻碍了在实际应用中的部署。为此,提出一种轻量级的密集残差卷积神经网络(dense residual convolutional neural networks,DRCNN)用于水下图像增强。为降低计算成本,DRCNN采用深度可分离卷积提取高级特征;通过密集连接和残差学习促进不同通道之间的信息交互,提高模型表征能力;将输入的退化图像与中间特征图融合,保留图像全局相似性,同时防止模型梯度消失。实验结果证明DRCNN能有效提高水下图像质量,较于现有算法,DRCNN参数量减少了85%,PSNR、SSIM值分别提高了3%、2%,测试速度提高了3%。DRCNN使用更少的参数实现了更好的性能,利于在低资源设备的实时场景中应用。
Deep convolutional neural networks are one of the main methods for underwater image enhancement,but their expensive memory consumption and computational requirements hinder their deployment in practical applications.To this end,a lightweight dense residual convolutional neural networks(DRCNN)is proposed for underwater image enhancement.DRCNN uses depthwise separable convolution to extract high-level features to reduce computational cost;promotes information interaction between different channels through dense connection and residual learning,but also improves model representation;and fuses the input degraded image with the intermediate feature map to preserve image global similarity while preventing model gradients from vanishing.The experimental results demonstrate that DRCNN can significantly improve the quality of underwater images.When compared to the existing algorithm,DRCNN parameters are reduced by 85%,PSNR and SSIM values are increased by 3%and 2%respectively,and test speed is improved by 3%.DRCNN achieves better performance with fewer parameters,which is advantageous for real-time applications on low-resource devices.
作者
周露跚
赵磊
李恒
刘辉
张国银
Zhou Lushan;Zhao Lei;Li Heng;Liu Hui;Zhang Guoyin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2023年第1期70-77,共8页
Journal of Electronic Measurement and Instrumentation
基金
国家自然科学基金(61863018)
云南省科技厅面上项目(202001AT070038)资助
关键词
水下图像增强
轻量级卷积神经网络
深度可分离卷积
密集连接
残差学习
underwater image enhancement
lightweight convolutional neural network
depthwise separable convolution
dense connection
residual learning
作者简介
周露跚,2020年于昆明理工大学获得学士学位,现为昆明理工大学的硕士研究生,主要研究方向为水下图像处理。E-mail:18468111430@163.com;通信作者:赵磊,2002年于昆明理工大学获得学士学位,2006年于昆明理工大学获得硕士学位,现为昆明理工大学实验师,主要研究方向为图像处理、机器学习。E-mail:zhaolei@sohu.com
