基于Mask R-CNN的防波堤复杂护面块体检测和分割方法

针对斜坡式防波堤护面层块体个数统计效率和精确率低的问题,提出了基于Mask R-CNN深度学习网络的斜坡式防波堤扭王字块体的识别和分割方法。该方法利用Mask R-CNN深度学习网络学习实验室采集图像的特征信息,通过调整交并比(IOU)阈值得到评价指标最好的模型,并将该模型应用于现场防波堤图像护面块体的识别和分割。测试结果表明,IOU取0.5时,目标分割的平均精确率为91.83%,平均召回率为92.94%；将训练得到的网络用于识别无人机航拍现场的防波堤图像,扭王字块识别率可达90.7%,且拍摄角度和高度对识别精度影响不大。Mask R-CNN深度学习网络可实现密集、复杂护面块体的准确识别,具有良好的移植性和通用性。

Identification and segmentation technology of complex armour blocks of rubble mound breakwater based on Mask R-CNN

In order to address the low statistical efficiency and accuracy of the number of armour unit blocks of slope breakwater, a method for the recognition and segmentation of the slope breakwater accropodes based the Mask R-CNN deep learning network was proposed. Firstly, this method used the Mask R-CNN network learning laboratory to collect the feature information of the image. Secondly, the model with the best performance evaluation index was obtained by adjusting the IOU threshold. Finally, the trained Mask R-CNN network was applied in the recognition and segmentation of the armour blocks of the on-site breakwater image. The test results show that when the IOU is 0.5, the average accuracy rate of target segmentation is 91.83% and the average recall rate is 92.94%. Using the trained model to detect the breakwater images taken by drones in actual projects, the identification rate of accropodes is 90.7%, and the shooting angle and height have little effect on the identification accuracy. Therefore, the Mask R-CNN deep learning network can realize the accurate recognition of dense and complex armour layer blocks with good portability and versatility.