In order to enhance the accuracy and robustness of the current crack detection
algorithm, we propose an improved algorithm based on FC-Densenet(Fully
Convolutional Densenet), named the Densely Connected Network with Attention
Mode(DCAN), for pixel-...
In order to enhance the accuracy and robustness of the current crack detection
algorithm, we propose an improved algorithm based on FC-Densenet(Fully
Convolutional Densenet), named the Densely Connected Network with Attention
Mode(DCAN), for pixel-level crack detection. We have incorporated SE
attention modules at five positions within the FC-Densenet framework.
Additionally, we have collected a private dataset specifically focusing on tiny
cracks, which closely resembles real-world crack scenarios. To validate the
effectiveness of our method, we conducted a series of experiments on three
publicly available crack datasets as well as our private dataset. Compared to the
baseline neural network, our proposed approach demonstrates superior
performance across six evaluation metrics. We observed that as the dataset size
increases, the advantages of our method become more pronounced in terms of
the mIoU and F1 metrics. For instance, on the Crack500 and Cross datasets,
our method achieved F1 scores of 83.24 and 80.40, respectively. Additionally,
the mIoU scores reached 71.11 and 68.80 on these two datasets, respectively.