Hierarchical Existential Prior Based on Expanded Pseudo Label for Crack Detection

Road crack detection approaches based on Image processing technique (IPT) have attracted much attention during the last decade for convenience and efficiency, but most of them can not achieve the expected performances due to the complex background interference and severe category imbalance of road images. This paper presents a hierarchical existential prior based on an expanded pseudo label for crack detection. Specifically, the framework contains three variants of UNet, and each sub-network is trained by pseudo labels generated by transforming semantic categories of non-crack pixels distributed in the neighborhoods of crack ones. Notably, the expansion degrees of labels for three sub-networks are set in hierarchical descending order. In other words, crack samples of the pseudo label for the latter sub-network are a subset of ones for the former one, and we define it as existential prior, which can optimize the network in a coarse-to-fine fashion and refine the detection result gradually. Besides, we utilize a hybrid loss consisting of IoU, SSIM, and Focal loss to optimize the network in different aspects, including image-aspect, patch-aspect, and pixel aspect in the training phase, which can improve the structural representation capability of the model. Additionally, we present a dynamic hyper-parameter adjustment strategy to balance the weight coefficients of different loss terms, which can enhance the robustness of the model for various practical scenes. Finally, the results on various datasets have demonstrated the effectiveness and superiority of the proposed method.