Deep Learning For Pavement Distress Detection

With the growing development of the national road networks,road maintenance is playing an essential role for keeping road safety and improving people’s livelihoods.Detecting pavement distress is an important step toward road maintenance since early locating and repairing the pavement distress can not only reduce the cost of pavement maintenance but also decrease the probability and severity of road accidents happening.Currently,the laser,radar,and other 3D scanners are ideal technologies for pavement distress data collection and the related algorithms are sophisticated.However,their implementation equipment is expensive and difficult to maintain,which is an obstacle to apply on a large scale.The method of pavement distress detection using a 2D camera mounted on the vehicle has the advantages of low cost,high flexibility,and easy installation.Whereas the accuracy of predictions cannot be guaranteed because the pavement distress does not have a certain shape and the appearance of the cracks usually changes drastically in different lighting conditions.Therefore,fully automated and comprehensive pavement distress detection is still challenging.The recent rise of artificial intelligence brings many possibilities in various tasks of computer vision,and several deep learning-based methods have made unprecedented progress in the field of pavement distress detection.However,these existing approaches still have some limits,such as low accuracy and high computational load.To overcome such limitations,extensive research has been carried out and the main work and contribution of this thesis are as follows:1.Pavement images are often captured in perspective view by the high-resolution camera that is installed in the car and far from the ground,causing the size of the pothole in the pavement image is relatively small.In this scene,the existing convolutional neural network is difficult to achieve a good balance between accuracy and efficiency.To address these issues,a location-aware convolutional neural networks is proposed,which focused on the discriminative regions in the pavement image instead of the global context.It consists of two main subnetworks: the first localization subnetwork(LCNN)employs a high-recall model to find as many candidate regions as possible in the downsampled image,and then only takes the top-ranked possibility of regions to get the highresolution parts through position remapping in the full-size image;the second partbased subnetwork(PCNN)eliminates some negative proposals while maintaining a high recall for the positives,and performs classification on the candidates on which the network is expected to focus.The experiments show that the location-aware convolutional neural networks can combine the advantages from both LCNN and PCNN to achieve high accuracy results while maintaining high computational efficiency.2.Cracks do not have a certain shape and the appearance of the cracks usually change drastically in different lighting conditions,making it hard to be detected by the algorithm with imagery analytics.Motivated by the recent success of the structures of semantic segmentation,an effective encoder-decoder fully convolutional neural network called UCrack Net is proposed.First,a dropout layer is added into the skip connection to achieve better generalization.Second,pooling indices is used to reduce the shift and distortion during the up-sampling process.Third,four atrous convolutions with different dilation rates are densely connected in the bridge block such that the receptive field of the network could cover each pixel of the whole image.In addition,multi-level fusion is introduced in the output stage to achieve better performance.Evaluation results demonstrate that the proposed method achieves high accuracy results,and outperforms other methods.3.A reliable and efficient pixel-level method of crack detection is critical for realtime pavement condition measurement.However,many existing encoder-decoder architectures for crack detection are time-consuming.According to the research and analysis on the structure of encoder-decoder architecture,a simple and effective method is proposed to boost the algorithmic efficiency based on encoder-decoder architecture for crack detection.It is able to dramatically speed up inference time while keeping the segmentation accuracy unchanged.The core idea of this method is a learnable automatic switch module(SWM),which predicts whether the image is positive or negative and then skips over the decoder to save computation time when it is predicted as negative.The classical U-Net and Deep Crack are chosen as examples of the encoderdecoder architectures to show how SWM is integrated into the architectures to reduce computation complexity.The experiment results demonstrate that the proposed method is correctness and effectiveness.4.In order to simultaneously detect multiple types of defects on the pavement,an effective and efficient pavement distress network(UPDNet)is proposed,which is combined with our current research mentioned above.Then,a pavement distress dataset with multi-defects classes is created based on images copied from the existing public dataset and the defects are annotated at the pixel level.The positive sample augmentation and training fine-tuning strategies are used to minimize the issues caused by class imbalance.Finally,a new measurement is proposed to evaluate the proposed method and the results show that the proposed method can identify crack and pothole accurately,which will provide a reference value for the practical application.