Small‐Sample‐Efficient Pavement Crack Detection Using Dynamic Sliding Window and Weighted Deep Forest

Pavement crack detection is essential for proactive road maintenance; however, existing deep learning methods are heavily reliant on large annotated datasets and GPU computing resources, which limits their deployment in resource‐constrained environments. To address this challenge, this study proposes a novel method called Dynamic Sliding Window and Weighted Deep Forest (DSW‐WDF) for efficient and accurate crack detection. The method integrates three core components: (1) Non‐local Means (NLM) filtering for preprocessing, which suppresses noise and uneven illumination while preserving crack edges; (2) a Window Change Value (WCV)‐driven dynamic sliding window that adaptively captures multi‐scale crack features, including ultra‐fine cracks measuring less than 2 mm; and (3) a weighted cascaded forest where subtree weights are optimized through cross‐validation to enhance decision reliability. Experimental results on a small‐sample dataset consisting of 3000 original images demonstrate that DSW‐WDF achieves an Intersection over Union (IoU) of 87.2%, surpassing traditional Deep Forest, ResNet50, and mainstream segmentation models such as UNet and DeepLab‐V3+. Notably, the model operates with only 710 MB of memory and achieves a real‐time inference speed of 28.5 FPS on a CPU, while reducing training time by 21% compared to traditional Deep Forest. This work validates the potential of non‐neural network ensemble models for small‐sample crack detection, offering a cost‐effective and easily deployable solution for road maintenance departments facing challenges related to data scarcity and limited computing power.