REDS-Net: uncertainty aware road crack segmentation based on residual enhanced Dilated segmentation network for ADAS

Road crack segmentation plays crucial role in ensuring road safety and efficiency in the movement of Advanced Driver Assistance Systems on the road. Traditional segmentation methods often suffer from challenges such as poor generalisation to unseen data and a lack of uncertainty estimation, which are critical for high-stakes applications like ADAS. This paper proposes a robust road crack segmentation framework, REDS-Net (Residual Enhanced Dilated Segmentation Network), that addresses these issues by integrating the stochastic dropout inference with a custom U-Net architecture. The proposed model incorporates the Dilated Depthwise Feature Extractor to capture fine-grained crack details, while the SDI enables uncertainty estimation by performing stochastic sampling during inference. The results demonstrate significant improvements in key metrics, including precision, recall, IoU, and F1-score, when compared to baseline methods. In addition, we adopt binary cross-entropy loss for the training of REDS-Net to compute the loss independently for each pixels and penalises incorrect predictions based on predicted probabilities. Experiments were conducted on six publicly available datasets: CFD, Crack500, CrackTree200, Deepcrack, GAPs384 and Forest. Furthermore, REDS-Net shows strong generalisation to unseen domains such as wet and uneven lighting road surfaces. The experimental results reveal that the proposed network outperforms the traditional road crack segmentation networks.