Effective Pavement Crack Delineation Using a Cascaded Dilation Module and Fully Convolutional Networks

Benkhoui, Yasmina; El-Korchi, Tahar; Reinhold, Ludwig

Crack detection in concrete surfaces is a critical structural health monitoring task. In fact, cracks are an early indication of the decaying of the structure that can lead to severe consequences. Manual inspection is time-consuming, costly, and contingent on the subjective judgment of inspectors. To address these challenges, we propose to use state-of-the-art techniques in computer vision to approach the crack delineation problem as a semantic segmentation task where pixels of the same class (background or crack) are clustered together. Our proposed method uses dilated convolution to enlarge the receptive field and preserve the spatial resolution. In this work, we present a fully convolutional network that consists of an encoder, a cascaded dilation module, and a decoder. While the encoder extracts the feature maps from input images, the cascaded dilation module aggregates multi-scale contextual information and finally, the decoder fuses low-level features, performs pixel-wise classification, restores the initial resolution of the images and subsequently outputs the segmentation results. Based on the same meta-architecture, we compare three different dilated encoder-decoder (DED) models: DED-VGG16, DED-ResNet18, and DED-InceptionV3. The three models have been trained and validated using a dataset comprised of 40000 images. For evaluation purposes, we used common performance metrics for semantic segmentation tasks: Precision, Recall, F1-score, IoU, and ROC curves. Our results show that DED-VGG16 achieved the highest accuracy (91.78%) and generated precise visual semantic segmentation results.