Pavement Crack Detection Based on Sparse AutoEncoder

Traditional pavement crack detection system can hardly detect cracks accurately due to the complicated background noises over the pavement surface. So a novel crack detection method based on sparse autoencoder was proposed. Firstly, an anisotropy detection algorithm was adopted to select the potential crack patches. Then the features of crack patches were extracted through sparse autoencoder and then trained by softmax to classify. Finally, benefited by the tensor voting based spatial enhancement, the cracks were extracted after noises-removing. Experimental results show that the proposed method can meet the requirement of crack detection. It is superior to other traditional methods with high accuracy and robustness.