摘要
针对水下可见光图像因水体散射、吸收导致的噪声干扰、纹理模糊、颜色失真问题,以及传统增强算法计算量大且基本耗时长的缺陷,文中提出一种基于轻量化多尺度卷积神经网络(CNN)的水下图像增强算法。该算法采用U-Net结构,融合浅层纹理特征与深层语义特征,有效恢复图像纹理细节与颜色信息;引入轻量化特征提取模块,在精简模型参数的同时加快网络收敛速度;在主干网络中嵌入多尺度金字塔池化,强化多尺度特征提取能力,弥补传统算法细节恢复不足的短板;采用L1损失与结构相似性损失联合优化,提升图像亮度与对比度的恢复效果。为满足工程应用低延时需求,算法经量化后部署于嵌入式平台,通过调用嵌入式神经网络处理器资源加速推理,在Atlas200IA2上的前向推理耗时仅28 ms。公开水下数据集的实验结果表明,该算法在测试集上的水下图像质量度量与水下彩色图像质量评估指标分别达到4.33和0.63,验证了其增强效果的有效性。
This paper proposed a lightweight multi-scale convolutional neural network(CNN)-based underwater image enhancement algorithm to address the problems of noise interference,texture blur,color distortion,and high computational complexity and long time consumption of traditional enhancement algorithms caused by water scattering and absorption in underwater visible light images.The U-Net structure was used,which combined shallow texture features with deep semantic features to effectively restore the texture and color information of the image.A lightweight feature extraction module was introduced,which not only simplified the model parameters but also accelerated the convergence speed of the network.A multi-scale pyramid pooling was embedded in the backbone network for extracting multi-scale features and compensating for the shortcomings of traditional algorithms in detail restoration.By combining L1 loss with structural similarity index loss for joint optimization,the restoration effect of image brightness and contrast was improved.To meet the low latency requirements of engineering applications,the algorithm was quantized and then deployed on an embedded platform.By calling the embedded neural processing unit resources to accelerate the inference process,the forward inference on the Atlas200IA2 took only 28 ms.Through experiments on publicly available underwater datasets,the algorithm proposed in this paper achieved an underwater image quality measure of 4.33 and the underwater color image quality evaluation index of 0.63,respectively,on the test set,demonstrating the effectiveness of the proposed enhancement algorithm.
作者
张俊
罗凡
袁政
ZHANG Jun;LUO Fan;YUAN Zheng(The 58th Research Institute of China Electronics Technology Group Corporation,Wuxi 214035,China;Wuhan Second Ship Design and Research Institute,Wuhan 430000,China)
出处
《水下无人系统学报》
2025年第6期1065-1073,共9页
Journal of Unmanned Undersea Systems
关键词
水下图像增强
多尺度卷积神经网络
轻量化
U-Net结构
边缘端部署
underwater image enhancement
multi-scale convolutional neural network
light weight
U-Net structure
edge deployment
