多级卷积神经网络的沥青路面裂缝图像层次化筛选

如何快速准确地识别与评估沥青路面裂缝病害，已成为路面养护和保障道路安全的重要任务之 一。实际采集路面图像中往往存在大量的非裂缝图像，在保证裂缝图像无漏筛的前提下，尽可能提高裂缝图像 的精确率与非裂缝图像的真负例率，则对于降低人工筛选的工作强度，以及后续裂缝自动分割与病害损坏程度 评估具有重要实际意义。故此，提出了一种多级卷积神经网络的沥青路面裂缝图像筛选方法，由训练、微调与 验证三阶段构成，利用微调集获得 softmax 层输入微调增量。为避免裂缝图像召回率增加与精确率下降的问题， 在对比不同卷积神经网络筛除的非裂缝图像异同基础上，采用改进 AlexNet 作为一级筛选网络，VGG16 或 ResNet50 作为二、三级筛选网络的层次化处理模型。对于含噪声及复杂路面图像测试集的实验结果表明，三级 层次化筛选模型能在 100%召回裂缝图像时，达到高的真负例率及准确率。与其他方法的对比实验表明，所提 方法可有效解决沥青路面裂缝图像漏筛问题，且具有更好的检测效果。

Multi-level convolutional neural network for asphalt pavement crack image hierarchical filtering

 The quick and accurate identification and evaluation of asphalt pavement crack disease has become one of the important tasks of pavement maintenance and road safety. There are a number of non-crack images in the actual collected pavement images. On the premise of ensuring that there is no missing filter in the crack image, it is of great practical significance to improve the precision of crack images and true negative rate of non-crack pavement images as high as possible, thus reducing the work intensity of manual filtering, as well as subsequent automatic crack segmentation and disease damage assessment. A multi-level convolutional neural network method for asphalt pavement image filtering was proposed, which consists of three stages, i.e, training, fine-tuning and validation. The input fine-tuning increment of softmax layer was obtained using fine-tuning set. In order to avoid the problem that the precision decreases when the recall of crack image increases, based on the comparison of the similarities and differences of non-crack images excluded by different convolutional neural networks, a hierarchical processing model was proposed, in which the improved AlexNet was employed as the first level filtering network and VGG16 or ResNet50 as the second or third level filtering network. The experimental results on noisy and complex road images show that the three-level hierarchical filtering model can achieve high true negative rate and high accuracy when recalling crack images 100%. Compared with other methods, the experimental results show that the proposed method can effectively solve the problem of missing filter in asphalt pavement crack image, and can produce a better detection effect.