Automated Structural Deficiency Detection: An Enterprise-Scale Computer Vision Approach

The structural integrity of physical infrastructure, particularly commercial and residential roofing, is critical to economic stability and public safety. Traditional inspection methodologies rely on manual, hazardous, and highly subjective human evaluations, which struggle to scale and often fail to detect early-stage material degradation. In alignment with emerging U.S. national priorities to accelerate the adoption of robust, reliable, and highly scalable artificial intelligence systems for infrastructure and security applications, this paper presents an enterprise-scale automated defect detection pipeline. Utilizing a deep convolutional architecture based on the ConvNeXt-Base backbone, the proposed model is trained on a rigorously curated dataset of over 125,000 field-captured images across 42 fine-grained structural deficiency classes. By implementing a dual-stage mitigation strategy for long-tailed class imbalance—combining inverse frequency sampling and label-smoothed cost-sensitive loss—the system demonstrates exceptional reliability in uncontrolled field environments. The model achieves an overall weighted F1-score of 0.53 across a massive 25,500-sample validation set, with critical failure categories such as structural punctures and debris accumulation exceeding an 0.85 F1-score. This research demonstrates the commercial viability of adapting advanced open-source AI architectures to solve national-scale infrastructure monitoring challenges.