AquaGestureSynth: Generating Realistic Underwater Hand Gesture Images with Gaussian-Poisson GAN

Abstract

Training robust underwater gesture recognition systems is limited by the scarcity and cost of annotated underwater data. Attempts to generate synthetic data by blending overwater images to introduce turbidity, color distortion, and illumination variability often fail to produce realistic images. This paper proposes AquaGestureSynth, a two-stage generative blending framework based on a Gaussian-Poisson Generative Adversarial Network (GP-GAN), to synthesize realistic underwater gesture images. The proposed method blends segmented overwater hand gestures with underwater backgrounds, introducing turbidity, light attenuation, and chromatic shifts while preserving gesture semantics. The first stage employs a Blending GAN for coarse realism, followed by the application of Gaussian-Poisson gradient-domain refinement within a Laplacian pyramid in the second stage. Further, a loss that combines style, content, histogram, and total variation losses to enhance texture consistency and visual realism has also been introduced. AquaGestureSynth achieves superior results across multiple metrics, with Structural Similarity Index of 0.91, Fr´echet Inception Distance (FID) of 24.8, and Learned Perceptual Image Patch Similarity (LPIPS) of 0.101, thereby outperforming Cycle-GAN and Poisson blending. Downstream gesture classification models trained on our synthetic data achieve up to 96% F1-score, demonstrating the model’s utility in augmenting underwater gesture datasets for human-robot interaction.