Research On Pavement Crack Detection Based On Adversarial Domain Adaptation

Cracks are the early manifestation of most pavement diseases.Accurate detection of cracks is of great significance for the prediction of pavement performance and the prevention and control of pavement diseases.In recent years,the powerful image representation capabilities of convolutional neural networks have prompted many scholars to apply this technique to detect cracks,and achieved good results.However,the differences in pavement materials,shooting conditions,and collection equipment of different pavement crack datasets make their data distributions different,which ultimately leads to the crack detection model trained by a dataset is not robust to the new dataset,detection performance tends to drop significantly.In order to improve the detection accuracy,the traditional solution is to select part of the images in the new dataset for pixel-by-pixel truth labeling,and then use them as training samples to retrain or fine-tune the model.However,the manual labeling of crack ground-truth is not only costly and time-consuming,it is difficult to keep up with the speed of road image acquisition.Therefore,how to improve the detection performance of the trained model on the new dataset when the new dataset has no ground-truth is the key problem of crack detection using convolutional neural networks.Unsupervised domain adaptation is an effective method to solve the problem of reduced performance of convolutional models caused by the different data distribution in the same task.It has been widely applied to cross-dataset segmentation of urban scenes images,and has achieved success.Compared with urban scene image segmentation,crack detection is a more elaborate task,and the change of crack in pavement image is more diverse and complex.Therefore,it is difficult to obtain good results by directly applying the domain adaptation method of urban scene images to pavement images.Based on this research status,this paper based on different pavement crack datasets,using adversarial training to design a universal unsupervised domain adaptation strategy for crack detection to improve the cross-domain detection performance of the crack detection model trained by a dataset,achieving unsupervised,high-precision cross-dataset detection of pavement cracks.The main contents and conclusions of this paper are as follows:(1)The construction of baseline crack detection network: In order to objectively evaluate the improvement of the domain adaptation strategy proposed in this paper to the crack detection network performance,a unified baseline crack detection network is needed.Since most crack detection networks are developed based on three general semantic segmentation networks,FCN,U-Net and Seg Net.Therefore,this paper learns and summarizes the network architectures of FCN,U-Net and Seg Net,and then compares the crack detection performance of these networks on the same crack dataset.The network with best performance is selected as the baseline network for domain adaptive research of crack detection.The experimental results show that the crack detection performance of U-Net is the best among these networks.(2)Research on cross-domain detection of pavement cracks based on adversarial domain adaptation: Based on U-Net as the baseline crack detection network,this paper uses adversarial learning to perform domain adaptation at feature level and network output level,and proposes a feature adaptation module and an output adaptation module to reduce differences in distribution between the model training set and new datasets.Specifically,the feature adaptation module aims to guide the encoder of the baseline network to learn domain-invariant feature representation,narrow the difference in feature distribution between the source domain(training set with ground-truth)and the target domain(new dataset without ground-truth),so that the encoder of the baseline network can extract more accurate crack features in the target domain.Considering that the features learned for crack detection are high dimensions and complexity,it is not possible to accurately align the source and target domains in such a feature space.In addition,the crack detection results of the network are also affected by the network decoder.Therefore,this paper carries out domain adaptation not only on feature level but also on output level.The goal of the output adaptation module is to capture the semantic diference between the source domain prediction map and the target domain prediction map,and then guide the baseline network to produce a prediction map on the target domain similar to the prediction map on the source domain by backpropagating the difference.In order to comprehensively evaluate the performance of the proposed domain adaptation strategy,this paper compares the cross-domain detection performance of the domain adaptation network and the baseline network in several cross-domain detection experiments of pavement cracks.In addition,this paper integrates two domain adaptation modules into the FCN architecture,and conducts a cross-domain detection experiment of pavement cracks again to prove the universality of domain adaptation strategy.The experimental results show that the domain adaptation strategy proposed in this paper can effectively reduce the distribution difference caused by different pavement materials,shooting conditions and collection equipment in the pavement crack detection task,improving the detection performance of the trained model on the new dataset when the new dataset does not have ground-truth.In addition,the proposed domain adaptation strategy is universal.