电子测量与仪器学报

麦仁贵,王骥.基于关键特征传递的KFT-GAN水下图像增强模型[J].电子测量与仪器学报,2026,40(2):76-85

基于关键特征传递的KFT-GAN水下图像增强模型

Underwater image enhancement model based on generativeadversarial network with key feature transfer

DOI：

中文关键词: KFT-GAN 关键特征传递模块深度可分离卷积混合损失函数

英文关键词:KFT-GAN key feature transfer module depthwise separable convolutions hybrid loss function

基金项目:广东省普通高校重点领域新一代信息技术专项（2020ZDZX3008）资助

作者	单位
麦仁贵	广东海洋大学数学与计算机学院湛江524088
王骥	2.广东海洋大学电子与信息工程学院湛江524088; 3.广东省智慧海洋传感网及其装备工程技术研究中心湛江524088

Author	Institution
Mai Rengui	College of Mathematics and Computer Science, Guangdong Ocean University, Zhanjiang 524088, China
Wang Ji	2.College of Electronics and Information Engineering, Guangdong Ocean University, Zhanjiang 524088, China； 3. Research Center of Guangdong Smart Oceans Sensor Networks and Equipment Engineering, Zhanjiang 524088, China

摘要点击次数: 366

全文下载次数: 206

中文摘要:

针对水体对光的吸收和散射等现象导致的水下图像出现色偏、模糊和细节丢失等问题，提出了基于关键特征传递的KFT-GAN 水下图像增强模型。设计了 KFT 关键特征传递模块，解决了色彩、边缘和纹理等关键特征在网络中高效传递的问题，并结合深度可分离卷积构建了轻量化的网络模型，与使用普通卷积构建的模型相比，参数量降低了60.6%，增强了模型的学习效率。同时该模块有利于生成器网络的编码阶段从输入图像中获取充足的关键特征，并通过下采样和跳跃连接将提取的关键特征传递至解码阶段进行图像重建，增强了重建图像的质量。另外，基于感知损失原理提出了混合损失函数，同时强调图像的多个关键属性，获得了视觉质量更好的图像效果。模型在EUVP和UIEB数据集上均取得了较好的性能，PSNR值分别为21.384 2和18.025 6，SSIM值分别为0.741 3和0.688 9，通过与传统算法和深度学习算法进行定性和定量的对比实验都证明了该模型的有效性和优越性。

英文摘要:

To address the challenges of color cast, blurriness, and detail loss in underwater images caused by light absorption and scattering, this study proposes a KFT-GAN for underwater image enhancement. The paper introduces a KFT module that facilitates the efficient transmission of critical features such as color, edges, and texture within the network. By integrating depthwise separable convolution, a lightweight network model is constructed, reducing the parameter count by 60.6% compared to models utilizing conventional convolution, thereby enhancing the model’s learning efficiency. This module enables the generator’s encoding stage to extract essential key features from the input image and transmit these features to the decoding stage through downsampling and skip connections, improving the quality of the reconstructed image. Additionally, this study proposes a hybrid loss function based on perceptual loss principles, emphasizing multiple key attributes of the image to achieve superior visual quality. The proposed model demonstrates excellent performance on both the EUVP and UIEB datasets, achieving PSNR values of 21.384 2 and 18.025 6, and SSIM values of 0.741 3 and 0.688 9, respectively. Qualitative and quantitative comparisons with traditional algorithms and deep learning methods validate the effectiveness and superiority of the proposed model.

查看全文查看/发表评论下载PDF阅读器