DPSO-NAS: Wall Crack Detection Algorithm Based on Particle Swarm Optimization NAS

Abstract

As urbanization progresses, building surfaces increasingly suffer from degradation and structural damage due to prolonged environmental stress, raising significant safety concerns. Consumer-grade drones with embedded vision technology offer a promising approach for intelligent detection of architectural surface anomalies. However, reliance on manually designed network architectures limits their effectiveness, as these struggle to represent complex textures, reduce crack segmentation accuracy, and fail to efficiently leverage the heterogeneous computing resources of drones, hindering widespread adoption in building inspections. To address these issues, we propose a Neural Architecture Search framework based on Dynamic Particle Swarm Optimization (DPSO-NAS). This framework introduces a hardware-aware search space to dynamically adapt architectures to drone computational constraints, a dual-path feature fusion unit using anisotropic convolution to enhance crack feature extraction, and an automated evaluation mechanism to eliminate human bias and ensure optimal model convergence. Experiments show DPSO-NAS outperforms manually designed networks by 4.7–12.3 percentage points in classification accuracy on CIFAR and ImageNet16-120 datasets. In crack segmentation, it achieves a 77.4% mIoU and reduces edge localization errors by 38.6%. On mainstream drone platforms, it improves inference speed by 2.1 times and cuts power consumption by 57%, advancing efficient, scalable inspection solutions.