融合深度学习与成像模型的水下图像增强算法

水下机器人的视觉感知功能因受到水下环境因素的影响，面临着图像质量降低的挑战，如图像颜色畸变、整体色调偏绿、偏蓝、对比度较低、细节较为模糊等。提出一种结合深度学习方法与物理成像模型的新型水下图像增强算法，通过构建包含扩张卷积和带参数激活函数的神经网络，进行背景散射光和直接传输映射的估计，并结合成像模型的数学表达进行重建运算得到增强后图像。实验结果表明，与UDCP、IBLA、GLNet等典型图像增强算法相比，该算法具有更快的运算速度，且能够消除水下环境因素带来的影响，丰富图像色彩的同时能增强各类细节，在峰值信噪比指标和结构相似度指标上取得了较大值。此外，增强后的图像在特征点匹配实验中获得了更好的匹配效果。

Underwater Image Enhancement Algorithm Combining Deep Learning and Image Formation Model

The visual perception of underwater robots is affected by environmental factors, which reduce the quality of generated images.Those images usually face color distortion, low contrast and loss of details, and the hue of the images tends to be green or blue.In this paper, a new underwater image enhancement algorithm that combines deep learning with an image formation model is proposed.The proposed algorithm estimates backscatter and direct transmission by building neural networks which include dilated convolution and parametric rectified linear units.Then the enhanced images are obtained by using mathematical expressions of the image formation model for reconstruction calculations.Experimental results show that compared with UDCP, IBLA, GLNet and other image enhancement algorithms, the proposed method displays a higher calculation speed.It can eliminate the influence of underwater environmental factors, enrich the color of images, and enhance details.It achieves higher scores in Peak Signal-to-Noise Ratio(PSNR) and Structural SIMilarity (SSIM) metrics, and the enhanced images provide better results in feature point matching experiments.