Deep convolution neural network model in problem of crack segmentation on asphalt images

Introduction. Early defect illumination (cracks, chips, etc.) in the high traffic load sections enables to reduce the risk under emergency conditions. Various photographic and video monitoring techniques are used in the pavement managing system. Manual evaluation and analysis of the data obtained may take unacceptably long time. Thus, it is necessary to improve the conditional assessment schemes of the monitor objects through the autovision.Materials and Methods. The authors have proposed a model of a deep convolution neural network for identifying defects on the road pavement images. The model is implemented as an optimized version of the most popular, at this time, fully convolution neural networks (FCNN). The teaching selection design and a two-stage network learning process considering the specifics of the problem being solved are shown. Keras and TensorFlow frameworks were used for the software implementation of the proposed architecture.Research Results. The application of the proposed architecture is effective even under the conditions of a limited amount of the source data. Fine precision is observed. The model can be used in various segmentation tasks. According to the metrics, FCNN shows the following defect identification results: IoU - 0.3488, Dice - 0.7381.Discussion and Conclusions. The results can be used in the monitoring, modeling and forecasting process of the road pavement wear.