A Joint Framework for Underwater Hyperspectral Image Restoration and Target Detection With Conditional Diffusion Model

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing 2024, Vol.17, p.17263-17277
• Main Authors: Li, Qi, Li, Jinghua, Li, Tong, Feng, Yan
• Format: Article
• Language: English
• Subjects: Deep learning; Diffusion; diffusion model; Diffusion models; Feature extraction; hyperspectral imagery (HSI); Hyperspectral imaging; Image degradation; Image enhancement; Image quality; Image restoration; Interference; Modules; Object detection; Performance evaluation; Reflectivity; Synthetic data; Target detection; Underwater; underwater image restoration; underwater target detection
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2024.3464557

Underwater hyperspectral imaging is crucial for various marine applications, with underwater hyperspectral target detection (HTD) holding significant importance. However, existing research on underwater HTD is limited, as most methods fail to adequately consider the impact of underwater target spectral variability and image quality degradation. To address these critical issues, we propose a novel joint framework for underwater hyperspectral image restoration and target detection, which is based on a conditional diffusion model. Our proposed framework consists of two main modules: the variable spectral group extraction module, and the joint underwater hyperspectral image restoration and target detection (JURTD) module. The variable spectral group extraction module leverages the conditional diffusion model to extract variable spectral image groups, thereby simulating the diverse range of underwater target spectra. Subsequently, the JURTD module extracts deep features from intrinsic images and the group of variable spectral images. Operating under the dual constraints of image restoration and target detection, this module achieves high-quality restored images and superior detection performance concurrently. Experimental evaluations conducted on both real-world and synthetic datasets demonstrate the effectiveness of our proposed framework in enhancing image quality and improving target detection performance. Moreover, the results indicate that our framework outperforms state-of-the-art methods, underscoring its practical utility and superiority in underwater hyperspectral imaging applications.