| 一种基于双通道的水下图像增强卷积神经网络 |
| |
| 引用本文： | 王树林,杨建民,卢昌宇,刘路平.一种基于双通道的水下图像增强卷积神经网络[J].海洋工程,2023,41(6):158-170. |
| |
| 作者姓名： | 王树林 杨建民 卢昌宇 刘路平 |
| |
| 作者单位： | 1.上海交通大学 海洋工程国家重点实验室，上海 200240
2.上海交通大学 三亚崖州湾深海科技研究院，海南 三亚 572024 |
| |
| 基金项目： | 海南省科技计划三亚崖州湾科技城自然科学基金联合资助项目（2021JJLH0001）；上海市科委资助项目（19DZ1207300） |
| |
| 摘 要： | 近年来各国对于海洋生物的保护意识日益强烈，用来监测海洋生物生存状态的水下机器人装备的研发是保护海洋生物资源的关键。水下相机是这类机器人在水下进行海洋生物监测时的光学感知设备。然而水下环境复杂，拍摄到的图像模糊不清，为解决水下图像模糊等问题，提出了一种基于双通道的水下图像增强卷积神经网络。在网络的编码器中采用双通道结构，其中一个通道采用了密集连接和高效通道注意力机制，提取水下图像的细节特征，另一个通道采用多尺度结构，提取原始图像的多尺度语义特征。接着，在网络中引入残差注意力模块和自适应特征融合模块，进一步优化了特征。最后将优化后的特征输入解码器重建出增强后的水下图像。试验表明：提出的网络算法在UIQM指标和Entropy指标上分别为3.005 6和7.654 7，较第二名的算法分别高出0.097 5和0.123 2。
|
| 关 键 词： | 卷积神经网络 水下图像增强 损失函数 密集连接 注意力机制 多尺度 |
| 收稿时间： | 2022/11/20 0:00:00 |
An enhanced convolutional neural network for underwater images based on dual channels |
| |
| WANG Shulin,YANG Jianmin,LU Changyu,LIU Luping.An enhanced convolutional neural network for underwater images based on dual channels[J].Ocean Engineering,2023,41(6):158-170. |
| |
| Authors: | WANG Shulin YANG Jianmin LU Changyu LIU Luping |
| |
| Abstract: | In recent years, countries have become increasingly aware of the protection of marine organisms. The research and development of robot equipment used to monitor the living state of marine organisms is the key to the protection of marine living resources. Underwater camera is an optical sensing device for marine biological monitoring. However, the underwater environment is complicated and the images taken are blurred. To solve the problems of underwater image blurring, a dual-channel enhanced convolutional neural network for underwater images is proposed. In the network encoder, a two-channel structure is adopted. One channel adopts the intensive connection and efficient channel attention mechanism to extract the detailed features of underwater images, while the other channel adopts the multi-scale structure to extract the multi-scale semantic features of the original images. Then, residual attention module and adaptive feature fusion module are introduced into the network to further optimize features. Finally, the enhanced underwater image is reconstructed by inputting the optimized features into the decoder. Experiments show that the proposed network algorithm has a UIQM index of 3.005 6 and an Entropy index of 7.654 7, which are 0.097 5 and 0.123 2 higher than the second algorithm. |
| |
| Keywords: | convolutional neural network underwater image enhancement loss function dense connection effective attention mechanism multi-scale |
|
| 点击此处可从《海洋工程》浏览原始摘要信息 |
| 点击此处可从《海洋工程》下载免费的PDF全文 |
