Severity Classification of Concrete Cracks Using Median Filtering and Deep Convolutional Neural Networks

Concrete crack severity (minor, moderate, severe) is a vital indicator for structural health monitoring and maintenance planning. Manual inspection methods, though common, are time-intensive, subjective, and often inconsistent. This study introduces a fully automated image-based approach combining grayscale conversion, median filtering, and a lightweight convolutional neural network (CNN) for classification of crack severity. From a large open-source dataset (Kaggle's Concrete Crack Images), we selected a balanced subset of ۹,۰۰۰ crack images (۳,۰۰۰ per category). Each image was resized to ۲۲۷×۲۲۷ pixels and preprocessed with grayscale conversion and a ۳×۳ median filter to suppress noise while preserving crack boundaries. A six-layer CNN trained on this dataset achieved ۹۳.۶% ± ۰.۶ accuracy on the filtered images, outperforming the raw images pipeline (۹۰.۸% ±۰.۹). Evaluation metrics including Precision, Recall, and F۱- score were also consistently above ۰.۹۲ for the filtered pipeline. The results indicate that this lightweight preprocessing improves classification performance, offering a practical approach for automated structural health monitoring (SHM) using visual data.

Keywords:

Concrete cracks , severity classification , preprocessing , CNN , structural health monitoring