DM-AECB: A Diffusion and Attention-Enhanced Transformer Framework for Underwater Image Restoration in Autonomous Marine Systems

Naveen Kumar Tiwari¹Dr.Abhishek Bajpai¹Shashank Yadav¹Dr Anas Bilal²Abdulbasit A. Darem³Raheem Sarwar^4*Jaibir Singh⁵

• ¹Rajkiya Engineering College Kannauj, Kannauj, India
• ²Hainan Normal University, Haikou, China
• ³Northern Border University, Arar, Saudi Arabia
• ⁴Manchester Metropolitan University, Manchester, United Kingdom
• ⁵Galgotias University, Greater Noida, India

Good and smart underwater vision is crucial for real-time observation of marine ecosystems as well as conservation, especially for autonomous vehicles such as autonomous underwater vehicles (AUVs) that traverse harsh oceanic conditions. In this paper, we introduce an innovative underwater image enhancement framework specifically designed for smart robotic systems involved in biodiversity monitoring, habitat mapping, and environmental sensing. Our approach integrates diffusion-based image restoration with an Attention-Enhanced Convolutional Blocks (AECB) enhanced Transformer backbone. A Denoising Diffusion Probabilistic Model (DDPM) allows for progressive refinement of the image, while the incorporation of AECB enables feature amplification selectively to enhance visual quality. To highlight distinctiveness from existing transformer-diffusion methods, our framework uniquely combines AECB modules for dual channel and spatial attention within the Transformer architecture. To facilitate on-board deployment in underwater AUVs and drones, this article presents a lightweight architecture and skip-sampling strategy that balance the performance of restoration and efficiency of computations. Research contributes to the creation of AI-fueled perception systems for intelligent ocean observation technologies to promote marine biodiversity protection. The proposed model and code will be made available on GitHub at: https://github.com/ntiwari91/DM-AECB.