Balancing Real and Synthetic Data for CNN-based Masonry Crack Detection

2026-06-06 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Intermediate, quick

Summary

Research on CNN-based masonry crack detection addresses the challenge of limited real-world data by exploring synthetic data generation. The study utilized a real dataset of masonry crack images from Bologna and a synthetic dataset created with a crack overlay tool. After identifying InceptionV4 as the best-performing model through initial training on real data, six training scenarios were tested, varying the ratio of real and synthetic data. Evaluation on a real image test set using F1-score and mean Intersection over Union (mIoU) metrics revealed that training with synthetic data plus a modest 20% real data achieved results comparable to, and in one case, outperformed, training solely on real data. Specifically, a 20% synthetic/80% real data scenario yielded a 76% F1-score and 80% mIoU.

Key takeaway

For Machine Learning Engineers developing CNN models for structural health monitoring, particularly masonry crack detection, you can significantly reduce the burden of collecting extensive real-world data. Consider integrating synthetically generated crack images, as a scenario with 20% real data combined with synthetic data achieved comparable or superior performance (76% F1-score, 80% mIoU) to using real data alone. This approach offers a practical path to robust models with fewer real-world data constraints.

Key insights

Synthetic data can significantly reduce real data collection needs for CNN-based crack detection while maintaining or improving accuracy.

Principles

• CNN performance depends on large, diverse datasets.
• Synthetic data effectively complements real-world data.
• Optimal data ratios enhance model accuracy.

Method

Train deep learning architectures on real data to select the best model (e.g., InceptionV4), then test varying real/synthetic data ratios, evaluating with F1-score and mIoU metrics.

In practice

• Generate synthetic cracks using overlay tools.
• Test a 20% real / 80% synthetic data mix.
• Evaluate crack detection with F1-score and mIoU.

Topics

• CNNs
• Synthetic Data
• Crack Detection
• Masonry Inspection
• Deep Learning
• Computer Vision

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Grounded in reality, new AI model spots fake images with less training — SimLBR detects fake images by learning real ones in latent space, offering computational efficiency and robustness.
• MS-SSE-Net: A Multi-Scale Spatial Squeeze-and-Excitation Network for Structural Damage Detection in Civil and Geotechnical Engineering — MS-SSE-Net enhances structural damage detection via multi-scale feature extraction and attention mechanisms.
• Rethinking Infrastructure Inspection as Image Difference Classification: A Traffic Sign Case Study — Image difference classification with instruction-based models reduces data dependency for infrastructure defect detection.
• Engineers issue hot take on cold-steel: Finding hidden damage requires radar, AI — A GPR and AI framework enables non-destructive detection of hidden damage in cold-formed steel structures.
• Synthetic Image Detection with CLIP: Understanding and Assessing Predictive Cues — CLIP-based synthetic image detectors rely on high-level photographic attributes, not just artifacts, but struggle with cross-generator generalization.
• SADGE: Structure and Appearance Domain Gap Estimation of Synthetic and Real Data — Synthetic data utility for computer vision tasks is best predicted by the non-linear interplay of appearance and geometric similarities, not either alone.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.