XCAT-UIR: explainable cross-attention guided ConvNeXt-transformer framework for underwater image restoration systems
Year: 2026
Type: article
Abstract: Underwater images play a significant role in ocean exploration, but they frequently suffer from major loss of image quality due to light absorption and scattering in underwater conditions. However, underwater image restoration remains a challenging issue because of various distortions. The primary cause of information degradation includes wavelength-dependent color attenuation, light scattering effects, and blurring of objects. Currently, deep learning methods are widely utilized for image enhancement, leveraging the concept that hidden features can be learned to improve the image quality. Nevertheless, the black-box nature of deep learning methods limits the understanding of how the model produces its restoration outputs. This study presents an Explainable Cross-Attention Guided ConvNeXt-Transformer for Underwater Image Restoration (XCAT-UIR) model. The major objective of the proposed XCAT-UIR approach is to increase the visibility, correct color distortions, preserve fine details, and improve the overall quality of degraded underwater images. The proposed approach initially performs preprocessing to improve image quality for further analysis. Subsequently, ConvNeXt and Swin Transformer techniques are employed to extract local texture and global contextual information from the preprocessed images to enhance the feature understanding. The integration of convolution-based and transformer-based feature extraction networks improves the representation capability of the restoration model. Besides, the Cross-Attention Fusion model is used to integrate local features from ConvNeXt and global features from Swin Transformer to generate discriminative fused feature representations. Furthermore, Residual Reconstruction Blocks reconstruct the enhanced underwater image from the fused features while preserving important details and reducing information loss. Lastly, Grad-CAM and cross-attention maps are used to enhance visual evidence for restoration-relevant areas and local-global feature interactions. To ensure the improved performance of the proposed model, a widespread experimental analysis is performed using a benchmark dataset. An extensive comparison study emphasized that the proposed model has gained significant performance over recent methods in terms of different measures, indicating its superiority in underwater image restoration.
Source: Scientific Reports
Institution: Multimedia University
Language: English
Topic: Image Enhancement Techniques
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7222843867