Road Crack Detection Based On Deep Learning

With the rapid development of road traffic,people’s demand for road maintenance is increasing.The traditional artificial pavement detection requires high personnel,time and safety,and is affected by subjective judgment.With the development of computer vision,people are constantly exploring the application of computer vision technology to road anomaly detection.The prediction results of automated road crack detection can be used as auxiliary materials to help inspectors detect.However,the training of crack detection task model can not be separated from a large number of rich data.Although the existing crack detection can improve the performance of the model through large-scale data collection or direct sharing of data centralized training model,data security is faced with huge risks and challenges in the process of data sharing and network transmission.In addition,large-scale crack detection data collection has the following problems: 1)Machine learning companies have strong data confidentiality,and private data cannot be disclosed;2)The use of multiple UAVs,smart cars and mobile terminals to collect cracks images may save time,while the pictures may involve privacy information;3)Single machine training requires huge time cost and data storage capacity.Based on the deep learning and federal learning methods,this paper models and analyzes the road cracks.The main research work and achievements are as follows:1.Due to the irregular shape and complex background of cracks,and the inaccurate location of small cracks and the fuzzy boundary of cracks exist in the current crack detection.Based on the characteristics of crack image data,combined with the powerful depth feature extraction ability of depth learning model,this paper constructs an effective neural network SFIAN to analyze the crack image data,selectively fuses multi-scale useful features at each stage,and models irregular crack targets,which improves the detection accuracy.Experiments on five data sets show that the model is effective and practical for road crack detection,and its detection accuracy F1 is about 13.3% higher than the baseline.2.Aiming at the problems in the data collection above,this paper proposes a lightweight crack detection model SFIAN+ based on SFIAN,which reduces the complexity of the model and reduces the memory consumption of the model.In addition,the model is trained by the federated learning algorithm with data protection,and finally a road crack detection model with good performance based on federated learning is obtained.Compared with SFIAN model,the memory usage of SFIAN+model parameters and models decreased by 67.6% and 67% respectively.In addition,in this paper the combination of federal learning and crack detection is studied for the first time,which achieves good prediction results and is expected to achieve good application in industry.