Pavement Crack Detection with Deep Learning Based on Attention Mechanism

To detect pavement crack automatically and accurately and improve detection effect, an attention-based crack network (ACNet) was proposed. ACNet adopts the encoder-decoder structure, and the encoder uses ResNet34 as the backbone network to extract pavement crack features. An attention-based feature module (AFM) is added between the encoder and the decoder, which utilizes the global information and increases the robustness of detecting different scales cracks, and it contributes to extract crack features and locate crack positions. The attention mechanism is also introduced in the decoding stage, and an attention-based decoder module (ADM) is designed to improve accuracy of detecting crack. Extensive experiments on two public crack datasets CFD and CRACK500 show that ACNet has better detection performances than other 8 methods, on the subjective vision, the crack location is more accurate and the details are more abundant; the experimental indicators F1 score and overlapping rate are significantly improved, which demonstrates the effectiveness of the proposed method. © 2020, Beijing China Science Journal Publishing Co. Ltd. All right reserved.