Automatic Crack Segmentation in Pavements using a Dilated Encoder-Decoder Network

Structural degradation of pavements has a direct impact on road safety. Due to weather conditions, moisture, and intensive use, pavements tend to crack which requires maintenance and repair. Over the last few years, crack detection has become an active research area with the advances of deep learning techniques. Nowadays, multiples techniques are being explored to automatically detect cracks in images. In this paper, we consider the crack detection problem as a semantic segmentation task where we differentiate between crack and noncrack pixels. To do so, we use an encoder-decoder architecture equipped with a dilated convolution module to better capture the contextual information and preserve the spatial resolution. The evaluation of the proposed architecture demonstrates its effectiveness at detecting cracks in pavements and achieves an mIoU of 81%.