Advancing Underwater Image Segmentation through Pix2Pix Generative Adversarial Networks

Underwater image segmentation presents significant challenges due to limited contrast, noise, light attenuation, and scarcity of labeled datasets. This paper proposes a novel application of Pix2Pix Generative Adversarial Networks (GANs) for multiclass underwater image segmentation. The proposed approach integrates a conditional GAN with a U-Net-based encoder-decoder architecture enhanced with skip connections and a PatchGAN discriminator to achieve pixel-level segmentation accuracy. The model was trained and evaluated on a custom dataset of 500 paired underwater images across five semantic categories: marine life, coral reefs, shipwrecks, rock formations, and seaweed. Our experimental results demonstrate superior performance compared to traditional methods, achieving an Intersection over Union (IoU) of 84.78%, Dice Coefficient of 90%, Precision of 84.2%, Recall of 86.4%, and F1-Score of 85.2%. The results validate the effectiveness of generative AI techniques in addressing the inherent challenges of underwater image analysis. This research contributes to advancing autonomous underwater vehicle (AUV) applications, marine environmental monitoring, and underwater exploration systems.

Keywords: Generative Adversarial Networks, Pix2Pix, Underwater Image Segmentation, Conditional GAN, U-Net, Image-to-Image Translation, Deep Learning