基于改进掩码-区域卷积神经网络的混凝土病害实例分割

Instance Segmentation of Concrete Defects Based on Improved Mask-RCNN

为对混凝土病害图像进行更精确的实例分割,提出改进掩码-区域卷积神经网络(Mask Region Convolution Neural Network,Mask-RCNN)。该网络采用轻量级的可移动网络(MobileNetV2)代替原始Mask-RCNN中卷积层过大的主干网络——残差网络(ResNet101),加入路径聚合网络(PANet),以提高Mask-RCNN提取浅层特征信息的能力。为验证改进Mask-RCNN的识别精度及其在实际工程中的可行性,首先构建多类混凝土病害图像数据集,利用K-means聚类算法确定最适合该数据集的先验边界框的长宽比,然后对比改进Mask-RCNN与原始Mask-RCNN、其它主流深度学习网络对混凝土五类病害(裂缝、露筋、剥落、白皙和空洞)的识别结果;最后利用无人机采集到的钢筋混凝土桥梁病害图像作为测试集进行测试。结果表明:改进Mask-RCNN在提高计算速度的同时能更准确地定位病害,减少了误检和漏检,识别精度高于原始Mask-RCNN及其它深度学习网络;改进Mask-RCNN可以识别无人机拍摄的未经训练的新的混凝土病害图像,识别精度满足实际工程需求。

In order to segment concrete defects images more accurately,an improved mask region convolution neural network(Mask-RCNN)is proposed.This improved Mask-RCNN uses MobileNetV2 to replace the residual network(ResNet101),both the backbone network of Mask R-CNN that has daunting convolution layers and the path aggregation network(PANet)were added to improve the ability of Mask-RCNN to extract shallow layer feature information.First,the multi-class concrete disease image data set was constructed,and using the K-means clustering algorithm to determine the aspect ratio of the most suitable prior bounding box for the data set.Then,the recognition results of improved Mask-RCNN,original Mask-RCNN and other mainstream deep learning networks on five types of concrete defects(cracks,exposed bars,spalling,efflorescence and voids)were compared.Finally,the defects images of reinforced concrete bridge collected by UAV were used as test set for testing.The results show that the improved Mask-RCNN can locate the defects more accurately,reduce the false detection and missing detection,and the recognition accuracy is higher than the original Mask-RCNN and other deep learning networks.The improved Mask-RCNN can identify new untrained concrete defects images taken by UAV,and the recognition accuracy can meet the practical engineering requirements.