Underwater Image Enhancement Using Generative Adversarial Network

The quality of the underwater image suffers due to water conditions like turbidity and light absorption. Underwater image enhancement is mainly to improve the visibility of the image and restores the colors and detail. Such enhancement is of considerable importance for scientific study and aesthetic quality in underwater images. This work proposed to use machine learning for improving the quality of the image. Generative Adversarial Network (GAN) model has been trained using a diverse dataset of underwater images to improve the color accuracy, contrast, and overall clarity. The resulting quality of images will be suitable for several applications such as in marine biology study, underwater archaeology, and navigation of autonomous underwater vehicles. The GAN model proposed includes generator and a discriminator. The generator accepts to the system a low quality underwater image and outputs as high quality image. Then the discriminator compares the generated image with the real high quality images, providing inner feedback into the generator. With this process the generator builds capacity to output realistic visuals. After the training phase, the model is tested on a set of test images, by measuring both numerical and visual parameters. Quantitative parameters including Brightness, Contrast, sharpness Saturation, noise Level, Dynamic range, PSNR were measured for evaluating the quality of the trained network.