Work

Fusion-based graph neural networks for synergistic underwater image enhancement

Year: 2024

Type: article

Abstract: Underwater images have become essential tools for marine exploration. However, their quality is often diminished by specific phenomena inherent to aquatic environments, thereby driving the need for research in Underwater Image Enhancement (UIE). Traditional UIE methods based on Convolutional Neural Networks (CNNs) offer advantages such as effective localized feature extraction but suffer from limited receptive fields that restrict their capability to capture long-range features, resulting in contextual and textural coherence issues. Graph Neural Networks (GNNs) excel in capturing long-range features due to their ability to process relational information across extended spatial contexts. Motivated by this strength , we have developed the Fusion-based Underwater Graph Network (FUGN), which initially segments images into blocks to transform them into a structure amenable to GNN processing. We employ Sobel and Gaussian Blur operators to compute similarities between these blocks, considering both texture and gradient information , which are crucial for constructing robust graph edges. The FUGN architecture synergistically combines the spatial precision of CNNs with the contextual depth of GNNs. This fusion harnesses the strengths of both neural network types, enhancing the overall capability of UIE tasks. Extensive experimental validations demonstrate the effectiveness of FUGN, with improvements of up to 6.49% in FSIM , 5.05% in PSNR , and 9.92% in SSIM, demonstrating marked improvements in the quality and fidelity of enhanced underwater images. Our code is available at https://github.com/zwh233/FUGN .

Source: Information Fusion

DOI: 10.1016/j.inffus.2024.102857