基于注意力的光照感知水下图像复原网络

Illumination Aware Underwater Image Restoration Network Based on Attention Mechanism

摘要:

水下环境光线昏暗，仅依靠自然光源难以清晰成像，通常需要增加人工光源，但人工光源的引入会导致场景亮度不均。在这种包含自然光源和人工光源的混合光照环境下，所拍摄的水下图像质量严重退化，不仅降低视觉观感，更影响后续高级计算机视觉任务的顺利开展。然而现有方法大都只考虑了自然光源的影响，对混合光源环境下的水下图像复原效果不佳。为了解决混合光源环境下水下图像存在的光照不均、 色偏、细节模糊等问题，提出了一个光照感知编解码器网络用于水下图像复原。一方面，在多尺度结构中引入注意力机制和改进残差结构高效提取丰富的结构细节特征，另一方面增加光照感知图作为先验约束网络复原结果的对比度。此外，设计了合适的损失函数，引导网络充分学习水下图像和清晰图像间的非线性映射关系，使恢复图像的色调更自然，纹理细节更丰富。对比试验结果证明此方法在主观感知和客观指标上均优于对比算法，消融实验证明所提网络模块和光照感知的有效性。

关键词:深度学习;光照感知;水下图像复原;注意力机制

Abstract:

The underwater environment is poorly lit，and it is difficult to image clearly with natural light sources alone. Artificial light sources are necessary to assist the illumination especially for underwater environment. However，inappropriate artificial lighting will lead to uneven brightness of scenes. The quality of the underwater images taken in this hybrid lighting environment with both natural and artificial light sources is severely degraded， which not only affects visual perception，but also poses a challenge to the successful execution of subsequent high-level computer vision tasks. However，most of the existing methods only consider the influence of natural light sources，and are not effective in recovering underwater images under hybrid light source environment. To handle the problems of uneven illumination，color bias and blurry details in underwater images under hybrid light source environments，we propose an Illumination-aware Encoder-Decoder Network(IEDN)for underwater image restoration. On one hand，attention mechanism and enhanced residual block are incorporated into a multi-scale structure to effectively extract detailed structure features. On the other hand，the illumination aware map is introduced as a prior constraint to balance the contrast of the restoration results. Meanwhile，appropriate loss functions are designed to guide the network to fully learn the nonlinear mapping relationship between the underwater image and the ground truth. In this way，the tone of the restored image is more natural and the texture details are more plentiful. The results of the comparison experiments prove that the proposed method is superior to other algorithms both quantitatively and qualitatively. The ablation experiments further demonstrate the effectiveness of the network modules and the illumination aware.

Key words:deep learning;illumination aware;underwater image restoration;attention mechanism