摘要
针对水下机器人在非限制环境中水下作业时,获取的水下图像存在整体色调偏蓝、偏绿、边缘细节较模糊及对比度较低等问题,提出一种基于灰度世界算法和端到端门控上下文聚合网络的水下图像增强算法。该算法通过图像R、G、B等3个通道分量调整的灰度世界算法,获取颜色校正后的水下图像;将校正后的水下图像输入到门控上下文聚合网络中,利用门控网络来融合图像中不同层次的特征,并引入平滑空洞技术和特征注意力模块,消除空洞卷积所出现的网格伪影现象,提高通道信息的灵活性,达到图像增强的效果。最后选取1 000幅参考图像,与6种经典增强算法进行主客观评价。结果表明,该算法在主观质量上提高了图像的对比度和清晰度,修正了水下图像的色偏;在客观评价指标上,测试集A中的峰值信噪比、结构相似性、信息熵和水下图像质量评估的平均值分别达到25.176 0 dB、0.950 9、8.057 9和0.618 2,测试集B的分别达到21.576 1 dB、0.933 1、8.119 4和0.591 4,评价结果都优于6种经典增强算法。
To address the overall bluish and greenish tones,blurred edge details and low contrast of underwater images received by underwater robots in unrestricted environments during underwater operations,an underwater image enhancement algorithm is proposed based on grey-scale world white balance and end-to-end gated contextual aggregation network.Firstly,the grey-scale world algorithm is designed to adjust the R,G and B components of the underwater image to obtain a colour-corrected underwater image.Secondly,the corrected underwater image is fed into a gated contextual aggregation network,and the gated network is used to fuse the features at different levels in the image,and the smooth cavity technique and feature attention module are introduced to eliminate the grid artefacts caused by the cavity convolution and improve the channel information flexibility to achieve the effect of image enhancement.Finally,1000 images with reference are selected and compared with six classical enhancement algorithms for subjective and objective evaluation.The results show that,the method proposed in this paper improves the contrast and sharpness of the underwater enhanced image in terms of subjective quality,and correct the color deviation of the underwater images.In terms of objective evaluation indicators,in test set A,the average values of peak signal to noise ratio(PSNR),structural similarity(SSIM),information entropy(IE)and underwater color image quality evaluation(UCIQE)reach 25.1760 dB,0.9509,8.0579 and 0.6182,respectively.In test set B,the average values of PSNR,SSIM,IE and UCIQE reach 21.5761 dB,0.9331,8.1194 and 0.5914,respectively.All of them achieve superior evaluation results to the six algorithms compared.
作者
刘真
高秀晶
洪汉池
LIU Zhen;GAO Xiujing;HONG Hanchi(School of Mechanical&Automotive Engineering,Xiamen University of Technology,Xiamen 361024,China;Institute of Smart Marine and Engineering,Fujian University of Technology,Fuzhou 350118,China)
出处
《厦门理工学院学报》
2024年第1期67-75,共9页
Journal of Xiamen University of Technology
基金
福建省海洋经济发展专项(FUHJF-L-2022-16)
福建省科技创新重点项目(2022G02008)
福建省财政厅教育和科研专项(GY-Z220233,GY-Z22011)。
关键词
水下图像增强
灰度世界
颜色校正
门控上下文聚合网络
特征注意力
underwater image enhancement
grayscale world
color correction
gated context aggregation network
feature attention
