A ConvNeXt-backbone cGAN for Underwater Power Engineering Image Enhancement

Underwater images is often greatly affected by color distortion, decreased contrast, and blurred details during the construction and maintenance of underwater power engineering. we proposed an underwater power engineering image enhancement model based on conditional generative adversarial networks (cGAN). Firstly, the generator incorporates the ConvNeXt module to enhance the global feature extraction capability by utilizing large convolutional kernels specifically designed for underwater image enhancement tasks. Subsequently, a boundary loss function is formulated to better capture high frequency information related to local textures and styles by comparing the boundary differences in the images. Lastly, a style transfer method is employed to transform land power engineering images and combine them with underwater environmental images to create a novel dataset for underwater power engineering. Case study results demonstrate the effectiveness of the proposed model in improving the quality of underwater power engineering images through adversarial learning, surpassing the enhancement performance of conventional models.