一种双注意力融合生成对抗网络的水下图像增强模型  

Dual attention fusion generative adversarial network for underwater image enhancement

作　　者：胡海峰 李凤英[1] HU Haifeng;LI Fengying(Guangxi Key Laboratory of Trusted Software,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学广西可信软件重点实验室,广西桂林541004

出　　处：《桂林电子科技大学学报》2023年第5期371-380,共10页Journal of Guilin University of Electronic Technology

基　　金：国家自然科学基金(62062029,61762024)。

摘　　要：针对水下图像在生成过程中会受到水下杂质污染以及光的吸收等问题,提出了一种双注意力融合生成对抗网络的水下图像增强模型。该模型使用了最新的Pix2Pix网络架构,并通过构建的双注意力机制结构建立丰富的上下文信息来处理水下图像,在模型生成器UNet网络首部增加了改进型Non-local模块,从多尺度角度获取更多全局特征,从而得到更加清晰的图像,在生成器尾部引入了Transformer模块,通过其优异的多头注意力块和多层感知机等结构来提升模型综合性能,从而进一步提升模型语义信息提取能力。实验结果表明,该模型在基准数据集EUVP上的峰值信噪比、结构相似性、水下图像质量评价指标相比其他模型平均提升了5.83%、4.88%和18.02%,而在基准数据集EUVP上的相应指标平均提升了6.21%、17.33%和15.96%。在主观可视化结果下,该模型也能适当处理图像退化问题,使图像呈现更好的清晰度和对比度。In order to solve the problem that underwater image will be polluted by underwater impurities and absorbed by light in the process of generation,an underwater image enhancement algorithm based on dual attention fusion generative adversarial network was proposed.The algorithm uses the latest Pix2Pix network architecture and constructs a dual attention mechanism structure to cre-ate rich context information to process underwater images.At the head of the model generator UNet network,an improved Non local module is added to obtain more global features from a multi-scale perspective,so as to get a clearer image.At the end of the genera-tor,a Transformer module is introduced to improve the comprehensive performance of the model through its excellent multi head at-tention block and multi-layer perceptron and other structures,so as to further improve the model Semantic information extraction ca-pability.The experimental results show that the peak signal-to-noise ratio,structural similarity,and underwater image quality evalua-tion indicators of the model on the benchmark dataset EUVP have increased by an average of 5.83%,4.88%,and 18.02%compared to other models,while the corresponding indicators on the benchmark dataset EUVP have increased by an average of 6.21%,17.33%,and 15.96%.Under the subjective visualization results,the model can also properly deal with the problem of image degra-dation,so that the image presents better clarity and contrast.

关 键 词：生成对抗网络 全局信息 改进型Non-local模块 双注意力融合 水下图像增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]