Abstract
Structural health monitoring (SHM) of civil, aerospace, and energy infrastructure increasingly relies on UAVs with vision sensors for efficient inspections. Crack classification is a central task, yet cloud-based inference introduces bandwidth, power, connectivity, and privacy challenges that limit its practicality. This study presents a fully self-contained Tiny Machine Learning (TinyML) pipeline for onboard crack classification on a milliwatt-level STM32H7 microcontroller. Using MobileNetV1x0.25 as the baseline, we systematically evaluate the full measurement pipeline, including image capture, preprocessing, and inference on a low-power embedded system. Two preprocessing strategies, a handcrafted sequence (grayscale, contrast, denoise, median, binarization) and a greedy algorithm-based composite method, are compared. Four compression techniques, namely post-training quantization (PTQ), quantization-aware training (QAT), pruning, and weight clustering, are assessed individually and in combination. The optimized pipeline achieves an F1-score of 0.938, an improvement of 11.4% over state-of-the-art deployments. At the same time, it requires only 2.9 MB RAM and 309 KB flash, with an end-to-end latency of 461.6 ms and an energy cost of 623.16 mJ per inference. On a DJI Mini 4 Pro UAV, continuous operation reduces flight time by just 1.31 minutes (4%), compared to 8 minutes (24%) when using Jetson-based platforms. Overall, this work delivers a reproducible benchmark for UAV-based SHM, demonstrating a practical balance of accuracy, resource efficiency, and energy consumption, and advancing the feasibility of on-device crack classification in highly resource-constrained environments.