High readability restoration of underwater-to-air text image for underwater devices based on URGAN

With advancements in maritime-to-aerial reconnaissance technologies and improvements in underwater devices, restoring distorted text images captured underwater for above-water situation awareness has become a key research focus. Conventional algorithms and deep learning-based methods often struggle to achieve clear and accurate text restoration. To address this challenge, a specialized dataset of underwater distorted text images was constructed using a large-scale scene text dataset and an underwater image distortion algorithm. URGAN (Underwater-text-image Restoration Generative Adversarial Network) is introduced as the first GAN-based method specifically designed for restoring underwater distorted text images. In particular, the generator of URGAN innovatively integrates numerous residual blocks and large convolutional kernels to preserve fine details. URGAN demonstrates strong performance in restoring text details and edges. In tests on simulated data, URGAN achieved a PSNR of 18.68 dB and an SSIM of 0.57. On real-world data, URGAN achieved a PSNR of 18.30 dB, an SSIM of 0.56, and a text recovery accuracy of 79.16%. These results confirm that URGAN generates highly readable restored images, showcasing its significant potential for applications in image processing for underwater devices.