基于物理先验的深度特征融合水下图像复原

Deep feature fusion for underwater-image restoration based on physical priors

由于水下环境的浮游生物悬浮杂质及不同光谱吸收率等干扰因素,水下图像往往会出现图像模糊、颜色失真和光照不均等退化问题。本文提出联合水下物理成像规律与数据驱动深度学习方法的水下图像重建模型。利用深度神经网络推断物理成像模型中的可学习参数,通过调制卷积和物理先验知识分别生成基于数据驱动的复原特征图和基于物理先验的复原特征图,引入混合注意力机制的深层特征级融合,重建最终的复原图像。实验结果表明该方法可以在减少噪声、提高对比度的同时,恢复图像的细节,提高水下图像的可视化质量和目标检测精度,增强水下学习模型的鲁棒性和泛化能力。

Due to interference factors such as suspended impurities of plankton and varying spectral absorption rates in an underwater environment,underwater images often suffer from degradation issues such as image blur,color distortion,and uneven illumination.This paper proposes an underwater-image reconstruction model that combines physical imaging principles with data-driven deep-learning methods.Using a deep neural network to infer the learnable parameters in the physical imaging model,the model generates data-driven restoration feature maps and physically informed restoration feature maps through modulated convolution and prior physical knowledge,respectively.Deep feature fusion with a mixed-attention mechanism is introduced to reconstruct the final image.Experimental results showed that this method can reduce noise,improve contrast,and restore image details,enhancing the visual quality and target detection accuracy of underwater images and increasing the robustness and generalizability of the underwater learning model.