CrackGA: Automatic Road Crack Detection Integrating Deep Learning and Genetic Algorithm

Crack detection in road infrastructure requires models capable of identifying fine, irregular patterns under varied and challenging conditions. Traditional U-Net architectures are effective for pixel-level segmentation, but manually designing an optimal structure for different crack types and environmental factors is time-consuming and demands domain expertise. In this work, we propose CrackGA, a framework that leverages Neural Architecture Search (NAS), guided by Genetic Algorithms (GAs), to automatically generate lightweight, task-specific U-Net variants. Rather than exploring the entire architectural design space, we define a domain-specific, block-wise search space tailored to the U-Net structure, balancing flexibility and computational efficiency. The genetic algorithm iteratively refines candidate architectures, optimizing the modular blocks of U-Net to explore diverse configurations. This approach not only automates architecture discovery for variable road conditions but also yields models with significantly reduced parameter counts and inference costs. Experimental results show that CrackGA achieves performance comparable to state-of-the-art models, with significantly fewer parameters, making it suitable for real-time deployment on edge devices.