| The transportation system is a large-scale public infrastructure and the security of it has significant influence on the operation of our society.Road cracks are among the commonest defects of the road system.And detecting cracks accurately is an effective way of monitoring the hidden dangers of the road system and ensuring the operation of the transportation system.Crack detection methods proposed recently mainly use traditional digital image processing algorithms.This kind of methods are normally slow and incapable of removing inference factors.In this paper we propose a road crack detection method based on optimized convolutional neural networks.Contributions made in this paper are as follows.(1)An optimized convolutional neural network structure using the idea of multi-scale images is proposed to improve the performance of CNN models in road crack detection tasks.This method fully uses the structure of multi-scale feature maps in CNN models,and makes up for the disadvantage of the high omission rate of linear CNN models using merely single-scale feature maps.(2)A weighting method for multi-scale feature maps using attention mechanism is used to further filter the feature maps in different layers.This part uses channel-wise attention mechanism to process the concatenated multi-scale feature maps to further select the multi-scale information,thus to improve the performance of feature extraction of CNN models.(3)A complete model of road crack detection system is proposed in this paper.This system combines traditional image processing methods with deep learning models,and is capable of accomplishing all of the procedures of image processing required to turn original images into detection outputs.Experiments show that the road crack detection system performs excellently in the real-world-gathered dataset as well as some public datasets,on which the precision of crack detection of our system is much higher than traditional digital image processing methods. |
PDF Full Text Request