A diffusion model and knowledge distillation framework for robust coral detection in complex underwater environments

Abstract

Coral reefs play a crucial role in marine ecosystems, but their sustainability is increasingly threatened by climate change and human activities. To aid in the protection and monitoring of these ecosystems, developing advanced artificial intelligence (AI)-based automated detection technologies is essential. This paper introduces the MambaCoral-Diffusion Detection framework (MambaCoral-Diffusion Detection, MCDD), an AI-driven approach for robust coral detection, designed to enhance performance in complex underwater environments—a critical challenge in marine engineering. Key AI contributions include integrating a diffusion model to generate realistic and diverse training data from limited and challenging underwater coral datasets, effectively alleviating the issue of data imbalance. Secondly, we adopted an innovative spatial sensing detection mechanism that enhances the accuracy of feature extraction in complex underwater environments. Finally, we introduced an efficient knowledge distillation technique that successfully transfers knowledge from complex models to more lightweight counterparts, thereby reducing computational resource requirements while maintaining efficiency and facilitating practical deployment. Experimental results show that MCDD achieves high performance on the Soft Coral dataset, reaching 91 frames per second (FPS), the mean average precision at 50% Intersection over Union (IoU) threshold of 0.843, and the mean average precision averaged across IoU thresholds from 50% to 95% of 0.566, with only 6.5 million parameters and 13.6 billion floating point operations per second (GFLOPs). These results demonstrate MCDD’s reliability and efficiency in detecting corals under complex underwater conditions, highlighting its significant potential for advancing marine research and conservation efforts. The code and dataset are available at https://github.com/RDXiaoLu/MambaCoral-DiffDet.git.