Research On Road Cover Defect Recognition Algorithm Based On Machine Learning

With the accelerating industrialization and urbanization in China,various municipal public underground pipeline facilities for drainage,gas,electricity and other cities are increasing.The correct detection and effective management of pavement manhole plays a key role in people’s personal safety.Manual inspection not only has disadvantages such as low efficiency,time consuming,and poor precision,but also other safety issues due to weather or the position distribution of the manhole cover.Therefore,this article has carried out a series of studies on how to realize the automatic detection of manhole cover.After studying and analyzing a large number of literatures on the detection of manhole cover defects,most scholars have arranged some hardware devices to obtain manhole cover information from the perspective of sensor and image processing.Finally,by analyzing these information,the test results are obtained.There are few methods for identifying and classifying manhole defects by combining video surveillance and machine learning algorithms.Although these methods can also realize the monitoring of the manhole cover,the performance is not good enough,and the calculation amount is large.It is not realistic to use it in the real-time system.Therefore,this paper proposes two different algorithms under machine learning to accurately and quickly realize the classification process of manhole cover defects.One is an improved SVM algorithm and the other is an optimized convolutional neural network.The main research contents are as follows:(1)By studying a large number of domestic and foreign literatures on SVM,feature extraction and incremental learning,the existing SVM algorithm has some defects in feature extraction of high-dimensional data.This paper proposes corresponding improvement methods.Firstly,the principal component analysis(PCA)and local linear embedding(LLE)methods are compared on the feature extraction method.It is found that the latter can better reduce the dimensionality of nonlinear data and preserve the essence of the data.The complexity is reduced on the basis of features,which greatly increases the speed of training and saves time costs.(2)After the feature extraction is completed,the classification is mainly realized by SVM.This paper proposes a support vector machine ASVM algorithm based on feedback mechanism and expert review.The algorithm will submit the SVM classification result to the expert for re-examination.If it is defect data,it will be put into the classified historical data set for secondary learning.Based on this,an incremental learning algorithm based on PAC-ASVM and LLE-ASVM is proposed to avoid the inefficiency of retraining all training sets when new data is added,and to classify handwritten data sets and manhole defects with a comparative test.(3)In addition to using support vector machine,this paper also proposes a fast and accurate classification of manhole cover based on optimized convolutional neural network.The optimization work is mainly reflected in the improvement of the activation function.For the gradient disappearance problem of Relu,the activation functions of MRelu and BRelu are proposed,and lots of comparative experiments were done on the public data set CIFAR-10.The experimental data proves that the proposed activation function not only improves the accuracy of the classification,but also speeds up the training of the model.In addition,considering the detection of the manhole cover requires a large number of manual labeling data sets,not only consumes a lot of manpower,but also inevitably causes errors,resulting in uneven labeling categories,which ultimately leads to inaccurate classifier identification.In this regard,an effective active learning method is proposed to solve the problem of sample labeling.