改进型VGG算法对小样本路面破损的分类识别

Classification Recognition of Pavement Disaster with Small Sample Size Based on Improved VGG Algorithm

为解决小样本沥青路面破损图片在分类识别中存在的识别精度差的问题,选择常见的5种路面破损类型,提出了一种基于VGG的浅层深度卷积神经网络模型的路面破损图像分类方法.首先,将采集到的图像集进行预处理并自制成数据集.其次,设置三种不同的批处理量尺寸和两种不同的网络层数作训练,选择最适合该网络模型的尺寸,得到浅层VGG模型.将处理后的路面图片直接作为模型的输入,作模型的训练、验证及测试.最后,将模型试验结果与支持向量机及目前主流的深度卷积神经网络模型的试验结果进行对比.结果表明:浅层VGG模型在训练集、验证集及测试集的分类准确率接近,对路面破损图像的分类识别准确率都达到98%以上,表现出模型良好的识别能力;与支持向量机及目前主流的网络模型试验结果相比,浅层VGG网络模型训练耗时少且泛化能力更强,模型提取到的特征更丰富,可获得更加全局的信息.可见,浅层VGG模型在对小规模图像的分类识别中具有显著优势,同时相比其他方法更具鲁棒性,结果更精确.

Aiming at the problems of poor identification accuracy in small sample size asphalt pavement damage classification recognition,five common types of pavement damage were selected.The shallow depth convolutional neural network model based on VGG is designed as an automatic classification method of asphalt pavement damage image.Firstly,the collected image samples are made into data sets for model training.Moreover,three different batch sizes and two different network layers are set up for training,and the most suitable size for the network model is selected so as to obtain the shallow VGG.The processed road image is directly used as the input of the model for training,verification and testing of the model.Finally,the test result was compared with SVM and the current main⁃stream deep convolutional neural network.The results show that the classification accuracy of the training set,verifi⁃cation set and a test set of shallow VGG is close,the classification and recognition accuracy rate of pavement damage image is more than 98%,which shows the good ability of recognition of shallow VGG.Compared with SVM and the current mainstream deep convolutional neural network,shallow VGG network model takes less time and has strong generalization ability,and can capture more global information.It can be seen that the shallow VGG model has sig⁃nificant advantages in the classification and recognition of small-scale images,it is more robust and the results are more accurate compared with other methods.