基于语义感知的多特征协同水下图像增强

Multi-feature Collaborative Underwater Image Enhancement Based on Semantic Perception

  • 摘要: 针对现有水下图像增强方法存在的去雾和颜色恢复效果不佳问题, 提出一种基于语义感知的多特征协同水下图像增强网络。首先, 通过语义感知特征提取模块充分提取原始水下图像的纹理和结构特征; 同时, 将水下退化图像输入到多尺度增强模块中进行对比度增强和多尺度特征细化, 并提取增强图像的全局特征; 然后, 在多特征融合模块中对多分支特征并行调制和融合, 增强多特征分支之间的有效联系和特征表达能力, 抑制冗余信息; 最后, 通过残差连接重构清晰的水下图像。在水下图像增强基准数据集、增强水下视觉感知数据集以及水下图像同步增强数据集上的实验结果表明, 所提方法相较基于像素处理及物理模型的经典算法、其他深度学习算法具有更好的增强效果, 全参考评价指标峰值信噪比及结构相似性分别提升1.72% 和4.3%, 无参考评价指标信息熵提升0.32%。

     

    Abstract: A semantic-aware multifeature collaborative underwater image enhancement network is proposed for existing underwater image enhancement methods with poor defogging and color recovery effects. First, the texture and structure features of the original underwater image are fully extracted using the semantic-aware feature extraction module. Meanwhile, the degraded underwater image is input to the multiscale enhancement module for contrast enhancement and multiscale feature refinement, and global features of the enhanced image are extracted. Second, the multifeature fusion module modulates and fuses multibranch features in parallel to enhance the effective connection and feature. Finally, a clear underwater image is reconstructed by connecting residuals. Experimental results on real-world underwater datasets UIEB, EUVP, and UFO show that the proposed method has a better enhancement effect than classical algorithms based on pixel processing, physical models, and other deep learning algorithms. In terms of objective evaluation indicators, the peak signal-to-noise ratio and structural similarity of full-reference evaluation indexes are improved by 1.72% and 4.3%, respectively, and the information entropy of the no-reference evaluation index is improved by 0.32%.

     

/

返回文章
返回