期刊
IEEE ROBOTICS AND AUTOMATION LETTERS
卷 6, 期 3, 页码 5121-5128出版社
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/LRA.2021.3070253
关键词
Image color analysis; Feature extraction; Absorption; Attenuation; Scattering; Iterative methods; Image enhancement; Marine robotics; underwater image enhancement; deep learning in robotics; recurrent network; computer vision for automation
类别
资金
- National Natural Science Foundation of China [61931022, 61671282]
- Open Fund of Key Laboratory of Advanced Display and System Applications of Ministry of Education (Shanghai University)
- Shanghai Science and Technology Innovation Plan [18010500200]
- Shanghai Shuguang Program [17SG37]
This study proposes a novel underwater image enhancement method by jointly optimizing the results of color correction and dehazing. It first uses a triplet-based color correction module to obtain color-balanced images, followed by a recurrent dehazing module to address the haze effect. Finally, an iterative mechanism is proposed to jointly optimize color correction and dehazing.
The captured underwater images suffer from color cast and haze effect caused by absorption and scattering. These interdependent phenomena jointly degrade images, resulting in failure of autonomous machines to recognize image contents. Most existing learning-based methods for underwater image enhancement (UIE) treat the degraded process as a whole and ignore the interaction between color correction and dehazing. Thus, they often obtain unnatural results. To this end, we propose a novel joint network to optimize the results of color correction and dehazing in multiple iterations. Firstly, a novel triplet-based color correction module is proposed to obtain color-balanced images with identical distribution of color channels. By means of inherent constraints of the triplet structure, the information of channel with less distortion is utilized to recover the information of other channels. Secondly, a recurrent dehazing module is designed to alleviate haze effect in images, where the Gated Recurrent Unit (GRU) as the memory module optimizes the results in multiple cycles to deal with severe underwater distortions. Finally, an iterative mechanism is proposed to jointly optimize the color correction and dehazing. By learning transform coefficients from dehazing features, color features and basic features of raw images are progressively refined, which maintains color balanced during the dehazing process and further improves clarity of images. Experimental results show that our network is superior to the existing state-of-the-art approaches for UIE and provides improved performance for underwater object detection.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据