Method of enhancement and coloring for underwater image based on multichannel image feature fusion

A method combining multi-source input fusion-based underwater image enhancement and GAN-driven grayscale image colorization is proposed to mitigate underwater imaging interference and lighten deep learning models for better generalization and real-time performance. Firstly, spatial domain image enhancement algorithms are utilized to pre-enhance the underwater images, obtaining predictions of red information in different channels. Then, the original images and pre-enhanced images are jointly used as input images, enabling the model to access more information. This model introduces the idea of fusing and reconstructing features from different levels of multiple input images, aiming to preserve as much original information and features as possible during the image enhancement process. Finally, a generator capable of colorizing grayscale images is trained using a large dataset of color images. The quality of the output colorized images is improved by defining an objective function composed of multiple loss functions. Experimental results show that compared to commonly used methods for low-light image enhancement and colorization, this method achieves better objective evaluation results in terms of peak signal-to-noise ratio, structural similarity, scale-invariant feature transform, thus verifying its excellent performance.