Research On Automatic Pavement Crack Detection Algorithm Fusing Deep Feature Aggregation Network And Attention Mechanism

Timely and accurate detection of pavement cracks is an important work to ensure road safety.Manual detection methods are time-consuming and highly susceptible to human factors.Although deep learning-based methods have made great progress in automatic crack classification and recognition,there are still difficulties such as poor anti-interference ability,large model parameters,and low detection efficiency.And it is difficult to achieve a good performance between detection accuracy and detection speed.For this reason,this work proposed an automatic pavement crack detection algorithm fusing deep feature aggregation network and attention mechanism.The specific works are as follows:(1)The unmanned aerial vehicle was used to collect 3000 real pavement crack images including 9 kinds of noise interferences,and then the crack images were enhanced and pixellevel marked.Finally,the overlapping sliding window technology was used to cut the crack images,so as to establish a real and reliable pavement crack dataset.(2)To deeply understand the principles and challenges of deep learning models in the crack detection,this work researched the pavement crack detection algorithm based on the Deep Labv3+.Next,this work effectively combined the Seg Net and the dense condition random field(DCRF),and proposed a pavement crack detection algorithm based on the Seg Net-DCRF.(3)To achieve a better result between crack detection accuracy and detection speed,and to improve the anti-interference ability of the model.This work proposed an automatic pavement crack detection algorithm fusing deep feature aggregation network and attention mechanism.This algorithm integrates the idea of deep feature aggregation network with spatialchannel squeeze and excitation module,which can make full use of the multi-scale receiving field to refine the crack detection results for many times.Meanwhile,the real-time reasoning ability of the model is improved.And the model contains multiple interrelated coding streams,which can fuse the high-level context information into coding features,thereby extracting the high-level features and low-level details of the crack.The attention mechanism can enhance the important information features,while suppressing the unimportant information features in the space and channels.Therefore,the accuracy of crack detection is effectively improved.(4)This work performed the morphological expansion and corrosion operations on the detected crack images,and then used the improved Zhang-Suen(ZS)refinement algorithm to extract the crack skeleton.Finally,the physical size of the crack was calculated according to the geometric characteristics of the crack,so as to comprehensively evaluate the health of the pavement.