Automated road maintenance depends on the accurate recognition of pavement distresses. Traditional manual inspection methods have difficulty meeting the high-frequency and high-precision detection for automated pavement maintenance. This paper introduces the junctional crossing net (JCNet) model, a dual encoding–decoding model capable of achieving pixel-level segmentation of multiple pavement distresses by fusing two-dimensional images and three-dimensional images. The proposed network uses a hybrid convolution module to fuse shallow and deep features to improve the delineation of pavement distress edges. A learnable split and reconstruction module is developed to capture the detailed information of the feature map. On the private asphalt pavement dataset and the public Crack500 benchmark, the experiment results show that the proposed model consistently outperforms the state-of-the-art deep learning models. The F-measure and mean intersection over union of JCNet on the private dataset are 95.92% and 92.76%, respectively. For complex environments such as water spot, foliage, and poor lighting conditions, the proposed JCNet model demonstrates notable noise immunity compared to other models.
Cheng et al. (Fri,) studied this question.