基于卷积自编码的沥青路面目标与裂缝智能识别

目前基于深度学习的路面裂缝识别经常面临训练数据集小，以及路面图片标注成本高等问题，基于小规模路面图片数据集，利用卷积自编码（CAE）方法进行数据增强，开展包括路面裂缝在内的路面目标智能化识别方法研究。在传统图像几何变换数据增强的基础上，采用CAE重构图片方法对原始数据集进行两步骤扩增；利用卷积神经网络DenseNet，设置了不同数据扩增方法的对比试验；针对沥青路面裂缝图片背景较黑，裂缝特征不清晰，无监督聚类学习难度大等问题，采用了一种基于CAE预训练的深度聚类算法DCEC，对经数据增强的路面图片进行无标注的聚类识别。研究结果表明：经过DenseNet网络100代的训练，在同一测试集的测试下，基于原始数据集训练的网络分类准确度为78.43%，利用传统图像处理方法进行扩增后准确度为83.44%，利用所提出的图片增强方法进行数据扩增后准确度达87.19%；在保持扩增后数据集样本量大小相同的情况下，与几何变换、像素颜色变换等经典数据增强手段相比，CAE重构图片的数据扩增方法有较高的路面图片识别精度；CAE数据扩增方法较受训练数据集样本量的影响，利用传统方法将数据集扩增后进行CAE特征学习，重构后的图片样本更易被机器识别；相较于传统机器学习聚类算法，所提出的的DCEC深度聚类方法将聚类准确率提升了约10%，初步实现了无需人工标注的路面目标的端到端智能识别。

Automatic Identification of Pavement Objects and Cracks Using the Convolutional Auto-encoder

The automatic detection of pavement cracks can significantly improve the efficiency of road maintenance for pavement engineers. At present, the Artificial Intelligence-based pavement crack detection may have the problems of insufficient training dataset or large dataset of pavement images that require high costs for manual classification and labeling. To solve these problems, based on the small-scale pavement image dataset, a study on the applications of the supervised and unsupervised deep learning models using the convolutional auto-encoder (CAE) method was conducted to identify different pavement objects, including the pavement cracks. Based on the traditional data augmentation using geometry transformation, comparison tests based on different data augmentation methods were conducted to validate the accuracy of the proposed research. Considering the problems that the background of the asphalt pavement crack image is dark, the crack characteristics are not clear, and the unsupervised clustering is difficult, a deep clustering algorithm DCEC (deep convolutional embedded clustering) based on CAE pre-training is proposed to study the road images. Test results show that:after 100 iterations of DenseNet network training, under the same test set, the test accuracy of network classification based on the original data set is 78.43%, the test accuracy based on the traditional data augmentation using image transformation method is 83.44%, and the test accuracy based on the method proposed in this study is 87.19%. It can also be found that, under the same dataset sample size, compared with the traditional data augmentation methods such as geometric transformation and pixel color transformation, the data augmentation method using CAE reconstruction has a higher recognition accuracy. Results show that the CAE data augmentation method is more easily affected by the quality and sample size of the training data set. After the data set is augmented by the traditional method, CAE learning is then carried out, and the reconstructed image sample is more easily recognized. Compared with the traditional method, the DCEC deep clustering method can improve the accuracy of clustering by about 10%, which preliminarily realizes the end-to-end intelligent recognition of road targets without manual annotation.