Pavement crack detection with multi-scale feature extraction and deep supervised feature fusion

In order to ensure the safety of the road surface, it is very important to detect the cracks in time and accurately. This paper proposes a pavement crack detection network model based on multi-scale feature extraction and deep supervised feature fusion. Firstly, multi-scale separable convolution blocks are used to extract crack features, which makes the model more effectively model pavement cracks with different topological structures. Then, the obtained feature maps are added to the deep supervised network to aggregate multi-level features, so that the model can converge faster and better. Finally, in order to make full use of the lateral output feature information of each stage, the dynamic feature fusion method is used to realize the efficient integration of multi-level features, which significantly optimizes the completeness and accuracy of the final prediction map. According to the evaluation results on DeepCrack, CFD and Crack500 public datasets, the proposed method shows better performance than other methods in crack detection.