Abstract:

Optical detection is an essential tool for underwater detection at short range， while underwater optical images are ineffective for underwater applications due to their low signal-to-noise ratio， low contrast and non-uniform illumination. A common method of underwater optical image processing is background foreground segmentation of the image. There are two main types of methods for performing image segmentation： traditional segmentation methods and deep learning methods. Traditional segmentation methods are susceptible to poor segmentation due to illumination and noise， while common deep learning methods are susceptible to training data limitations and poor generalization. In this paper， a neural network structure with parallel convolution and a constrained loss function are designed， and the optimal values of the hyper-parameters of the loss function are obtained through extensive experiments. The results show that the MAE values of the experimental results obtained by this paper are much smaller than those of FCN8 and UNet， and the mIoU values are larger than those of FCN8 and Unet， and the P-R curves are better than those of other methods， which are more adaptable to the complex and variable characteristics of the underwater environment and can obtain better segmentation results.

Key words: threshold segmentation, underwater image, non-uniform illumination, parallel convolution, deep learning

摘要：

光电探测是水下目标近距离探测时不可缺少的技术手段。水下光学图像具有低信噪比和对比度、照明不均匀等特点，使得光电探测在水下应用时效果不佳。常见的水下光学图像处理方法是对图像进行背景前景分割。目前主要有2种分割方法，传统分割方法易受光照、噪声等因素影响，效果不佳；深度学习方法易受训练数据限制，泛化能力不强。设计了一个带并行卷积的神经网络结构以及带约束的损失函数，通过大量实验获得了损失函数的超参数最优取值，并在不同照明条件、不同浑浊度、光照不均匀的条件下进行了实验分析。结果表明：该方法实验所获的MAE值远小于FCN8，UNet等方法，mIoU值大于FCN8，Unet等方法，P-R曲线优于其他方法的P-R曲线，更加适应水下环境复杂多变的特点，可获得更好的目标分割结果。

关键词: 阈值分割, 水下图像, 光照不均, 并行卷积, 深度学习