DM-AECB: a diffusion and attention-enhanced convolutional block for underwater image restoration in autonomous marine systems

Abstract

Introduction Effective underwater vision is critical for real-time marine ecosystem observation and conservation, especially for autonomous underwater vehicles (AUVs) operating in challenging oceanic environments. Methods We propose a novel underwater image enhancement framework tailored for smart robotic systems used in biodiversity monitoring, habitat mapping, and environmental sensing. Our method integrates a Denoising Diffusion Probabilistic Model (DDPM) for progressive image restoration with an Attention-Enhanced Convolutional Blocks (AECB) augmented Transformer backbone. The AECB modules provide dual channel and spatial attention, selectively amplifying features to enhance visual quality. Additionally, a lightweight architecture combined with a skip-sampling strategy is designed to optimize computational efficiency for onboard deployment in AUVs and underwater drones. Results Experimental evaluations demonstrate that our framework achieves superior image restoration performance while maintaining computational efficiency, outperforming existing transformer-diffusion approaches. The dual attention mechanism within AECB modules distinctly improves the clarity and detail of underwater images. Discussion This work advances AI-driven perception systems for intelligent ocean observation technologies, supporting improved marine biodiversity protection. The proposed model promises practical real-time applications in autonomous underwater exploration and monitoring. The model and code will be made publicly available on GitHub: https://github.com/ntiwari91/DM-AECB .