采用深度卷积神经网络的路面破损智能识别

为有效识别沥青路面病害类别,为后续养护对策的制定提供依据,将深度卷积神经网络,视觉几何组(Visual Geometry Group NetWork, VGG)技术引入沥青路面病害识别任务中。根据VGG网络结构随着卷积核深度的加深可获得图片更深层次特征的特点,将VGG模型最后一层卷积核中的卷积深度加深,获得改进后的VGG模型,并与VGG模型进行比较。结果表明：改进后的VGG模型用时为278ms/step,相比于VGG模型用时为258ms/step略有增加,而对病害的识别精度又进一步优化,提升了1.36,对龟裂、松散这类复杂裂缝分别提高了1.12%、8.4%。可见,采用VGG模型可以有效识别路面病害,将其适当改进后,效果更佳,相比于其他方法,对诸如松散、龟裂等复杂路面病害可做到精确识别,达到及时、有效监测、养护路面,防止路面进一步退化的目的。

Intelligent recognition of pavement damage using deep convolutional neural network

In order to effectively identify the types of asphalt pavement disease and provide the basis for the subsequent maintenance countermeasures, the deep convolutional neural network, Visual Geometry Group Network (VGG) was introduced into the asphalt pavement disease identification task. According to the feature that VGG model can extract the deeper features of the image as the convolutional kernel depth is deepened, the convolutional kernel depth in the last convolutional layer of the VGG model was deepened by experiment comparison, and the improved VGG model was obtained. The recognition accuracy of cracks were more than 97%, the accuracy of tortoise can up to 95.33%, that of loose cracks and pit slot is 96.41% and 97.08%, respectively. The improved VGG model uses 278ms / step, which slightly increases the time of 258ms / step, and further optimizes the disease identification accuracy, which increases by 1.36 percentage points, and increases by 1.12%, 8.4% respectively for tortoise, loose. It can be seen that the improved VGG model has more advantages. Compared with other methods, it can accurately identify the complex pavement disaster such as loose and tortoise and so on, which can achieve the purpose of timely and effective monitoring, maintenance of the pavement and prevent the further deterioration of the pavement.