摘要： 针对水下图像视觉质量退化以及单一卷积层特征利用率不高的问题，提出了一种基于多尺度特征提取的水下图像增强模型。本模型在卷积神经网络的基础上使用改进的多尺度特征提取模块提取水下图像特征，首先构建可分离残差密集块（SRDB）作为基本特征提取单元，使用SRDB模块进行残差密集连接得到多层次特征信息，最后融合3个不同初始感受野下的多层次特征信息作为该模块的输出。多组试验结果显示，本模型增强后的水下图像有效改善了颜色失真和低对比度现象的同时保持了丰富的边缘细节内容；EUVP测试集的PSNR、SSIM分别上升到28.52、0.88，真实河豚图像测试集的UIQM、NIQE分别上升到2.84、5.95，表现均优于对比方法。研究表明，本模型具有较高的 FPS，大幅提升水下图像视觉感知质量的同时保持了良好的实时性。

关键词: 水下图像增强, 多尺度特征, 残差连接, 感受野, 实时性 

Abstract: Aiming at the degradation of the visual quality of underwater images and the low feature utilization of a single convolutional layer, an underwater image enhancement model based on multi-scale feature extraction is proposed. The model uses the improved multi-scale feature extraction module to extract underwater image features based on the convolutional neural network. First, a separable residual dense block (SRDB) is constructed as the basic feature extraction unit, and the SRDB module is used for residual dense connection to obtain multi-level feature information, and finally merge the multi-level feature information under three different initial receptive fields as the output of this module. Multiple sets of experimental results show that the enhanced underwater image of the model effectively improves the color distortion and low contrast while maintaining rich edge details; The PSNR and SSIM on the EUVP test set rose to 28.52 and 0.88, respectively, and the UIQM and NIQE on the real pufferfish image test set rose to 2.84 and 5.95, respectively, outperforming the comparison methods. Research shows that this model has a high FPS, which greatly improves the visual perception quality of underwater images while maintaining good real-time performance.

Key words: underwater image enhancement, multi-scale features, residual connection, receptive field, real-time