基于改进U-Net的水草图像分割方法

无人艇（USV）在河道水面作业过程中，水草会缠绕推进器，这是整个业界应用都遇到的困扰。针对水面图像中水草分布的全局性、分散性以及边缘和纹理的复杂性，对U-Net进行改进并用于对图像所有的像素进行分类，以减少网络特征信息的丢失，并加强全局和局部特征的提取，从而提高分割性能。首先，采集多地多时段水草图像数据，制作了一个比较全面的水草语义分割数据集；其次，提出在U-Net中引入三个尺度的图像输入，从而使得网络对特征进行充分提取，并引进三种上采样图像的损失函数来平衡三种尺度的输入图像带来的总体损失；此外，还提出了一种混合注意力模块并引入到网络中，其包含空洞卷积和通道注意增强两个分支；最后，在新构建的水草数据集上对所提网络进行验证。实验结果显示，所提方法的准确率、均交并比（mIoU）和平均像素精度（mPA）值分别可达96.8%、91.22%和95.29%，与U-Net（VGG16）分割方法相比，分别提高了4.62个百分点、3.87个百分点和3.12个百分点。所提方法可应用于水面无人艇对水草的检测，并进行相应的路径规划来实现水草避让。

Waterweed image segmentation method based on improved U-Net

During the operation of the Unmanned Surface Vehicles （USVs）， the propellers are easily gotten entangled by waterweeds， which is a problem encountered by the whole industry. Concerning the global distribution， dispersivity， and complexity of the edge and texture of waterweeds in the water surface images， the U-Net was improved and used to classify all pixels in the image， in order to reduce the feature loss of the network， and enhance the extraction of both global and local features， thereby improving the overall segmentation performance. Firstly， the image data of waterweeds in multiple locations and multiple periods were collected， and a comprehensive dataset of waterweeds for semantic segmentation was built. Secondly， three scales of input images were introduced into the network to enable full extraction of the features via the network， and three loss functions for the upsampled images were introduced to balance the overall loss brought by the three different scales of input images. In addition， a hybrid attention module， including the dilated convolution branch and the channel attention enhancement branch， was proposed and introduced to the network. Finally， the proposed network was verified on the newly built waterweed dataset. Experimental results show that the accuracy， mean Intersection over Union （mIoU） and mean Pixel Accuracy （mPA） values of the proposed method can reach 96.8%， 91.22% and 95.29%， respectively， which are improved by 4.62 percentage points， 3.87 percentage points and 3.12 percentage points compared with those of U-Net （VGG16） segmentation method. The proposed method can be applied to unmanned surface vehicles for detection of waterweeds， and perform the corresponding path planning to realize waterweed avoidance.