Abstract: To address problems caused by deep-sea imaging, such as color distortion, low contrast, and blurred structures in underwater images, a Multiscale Parallel Enhancement Network (MPENet) was proposed from the perspectives of spatial, channel, and three-dimensional features. Image information was fed in parallel into a multi-dimensional feature extraction network and an encoder. A multiscale feature refinement module was introduced into the feature extraction network to further process extracted features, allowing the network to learn information at different scales more accurately. A multi-dimensional color enhancement module was incorporated into the encoder to enhance image details and colors. An adaptive enhancement network was then designed to further refine feature representations and fuse multi-level features. Finally, a decoder was used to generate the enhanced image. The proposed method demonstrates outstanding performance on public datasets. It achieves a peak signal-to-noise ratio (PSNR) of up to 24.8651 and a structural similarity index measure (SSIM) of 0.8954, representing improvements of 1.5806 and 0.0398 over HFM, respectively. Underwater color image quality evaluation (UCIQE) and underwater image quality measure (UIQM) reach 0.5931 and 3.1028, surpassing HFM by 0.0384 and 0.1514, respectively. Experimental results show that MPENet effectively improves underwater visual quality.

Key words: image processing, feature extraction, multiscale feature, deep convolution, intensified color

摘要： 针对深海拍摄会导致水下图像色彩偏移、对比度过低和结构不清晰等问题，结合空间、通道和三维特征的角度提出一种多尺度并行增强型网络(MPENet)。MPENet是将图像信息并行传入多维特征提取网络和编码器中。在多维特征提取网络中引入多尺度特征精炼模块进一步处理提取到的特征信息，使网络更准确学习不同尺度的信息。然后在编码器中引入多维色彩增强模块，增强图像细节色彩。最后设计自适应增强网络对特征信息进一步处理并融合多级信息，再通过解码器得到最后图像。此外，在公开数据集上所提算法表现优异，峰值信噪比(PSNR)和结构相似性度(SSIM)最高分别达到24.8651和0.8954，相比HFM算法提升了1.5806和0.0398。水下色彩质量评价(UCIQE)和水下图像质量测量(UIQM)分别达到0.5931和3.1028，相比HFM算法提升了0.0384和0.1514。实验结果表明MPENet能有效提升水下视觉效果。

关键词: 图像处理, 特征提取, 多尺度特征, 深层卷积, 强化色彩