A multi-color and multistage collaborative network guided by refined transmission prior for underwater image enhancement

• Published in: The Visual computer Vol. 40; no. 11; pp. 7905 - 7923
• Main Authors: Ouyang, Ting, Zhang, Yongjun, Zhao, Haoliang, Cui, Zhongwei, Yang, Yitong, Xu, Yujie
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.11.2024; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Collaboration; Color; Computer Graphics; Computer Science; Controllability; Deep learning; Deviation; Image contrast; Image enhancement; Image Processing and Computer Vision; Image quality; Image transmission; Low pass filters; Machine learning; Original Article; Parameter estimation; Saturation (color); Underwater; Physical-inconsistency loss; Underwater image enhancement; Transmission prior derivation; Multistage and multi-color space collaborative network
• ISSN: 0178-2789; 1432-2315
• DOI: 10.1007/s00371-023-03215-z

Due to the attenuation and scattering properties of light in underwater scenes, underwater images are generally subject to color deviations and low contrast, which is not conducive to the follow-up algorithms. To alleviate these two problems, we propose a multi-color and multistage collaborative network guided by refined transmission, called MMCGT, to accomplish the enhancement tasks. Specifically, we first design an accurate method of parameter estimation to derive transmission priors that are more suitable for underwater imaging, such as min–max conversion, low-pass filter-based estimation and saturation detection. Then, we propose a multistage and multi-color space collaborative network to decompose the underwater image enhancement task into more straightforward and controllable subtasks, including colorful feature extraction, color deviation detection, and image position information retention. Finally, we apply the derived transmission prior to the transmission-guided block of the network and effectively combine the well-designed physical-inconsistency loss with Charbonnier loss and VGG loss to guide the MMCGT to compensate for the quality-degraded regions better. Extensive experiments show that MMCGT achieves better evaluation results under the dual guidance of physics and deep learning than the competing methods in visual quality and quantitative metrics.