Roadway Crack Segmentation Based On An Encoder-Decoder Deep Network With Multi-Scale Convolutional Blocks

In highway pavement management, to detect and segment cracks are the key distresses in the condition evaluation. Considering the characteristics of imbalance, noise corruption and various sizes of cracks samples, we propose a novel fully convolutional architecture to improve the efficiency and accuracy of crack segmentation. This architecture adopts an encoder-decoder network, to make full use of information characteristics through skip connections from different dimensions by merging features at various levels. Meanwhile, in order to obtain feature maps with different reception field, we use different sizes of convolution kernels and concatenate generated feature maps, so that not only cracks of different sizes are detected but also noise are suppressed. To solve the problem of sample imbalance, we design the loss function by combining the DICE coefficient with binary cross entropy to improve the performance of segmentation. We train and evaluate our architecture to public crack data sets AigleRN and CFD, for pixel-wise prediction and obtain excellent performance compared with other methods.