基于红色通道注意力机制的水下图像增强

Underwater image enhancement based on the red channel attention

摘要:

水下图像增强因其在海洋勘测和水下机器人中的重要意义而备受关注。在过去的几年中，已经提出了许多水下图像增强算法。已有的深度学习方法由于忽略水下图像的预处理过程和对红色通道信息的增强或者弱化了这个过程导致增强结果并不显著，其往往只适应特定的场景，缺乏泛化能力。为此，我们基于卷积神经网络建立了一种全新的水下图像增强算法，为了充分利用特征图的通道信息，在相同维度的特征图之间采用不同尺寸的卷积核获取更多通道数目的特征。然后，我们基于红色通道构建了注意力机制，以加强对于图像中容易丢失信息的红色通道的特征的提取。最后，我们在EUVP,UFO120数据集做了消融实验证明了红色通道注意力机制的有效性。通过对对比实验的增强结果进行各项指标分析，证明增强结果有着更高的的结构相似性和峰值信噪比，并且在无参考指标方面有着更高的颜色平衡、清晰度以及对比度，综合性能优于以往的方法。

Abstract:

Underwater image enhancement has attracted attention for its importance in marine survey and underwater robotics. In the past few years, many underwater image enhancement algorithms have been proposed. The existing deep learning methods ignore the pre-processing process of underwater image and the enhancement of red channel information, or weaken the enhancement results, which often only adapt to specific scenes and lack the generalization ability. To this end, we establish a brand new underwater image enhancement algorithm based on convolutional neural network. In order to make full use of the channel information of feature maps, convolutional cores of different dimensions are used to obtain more number of channels in the same dimensions. Then, We constructed an attention mechanism based on red channels to enhance the extraction of features for red channels that tend to lose information in the image. Finally, we did ablation experiments in the EUVP, UFO120 dataset to demonstrate the effectiveness of the red channel attention mechanism By analyzing the various indicators of the contrast experiment enhancement results, we prove that the enhancement results have higher structural similarity and peak signal-to-noise ratio, and have higher color balance, clarity and contrast in terms of reference-free indicators, and the comprehensive performance is better than the previous methods.