An automated classification and evaluation tool utilizing a convolutional neural network has achieved an overall classification accuracy of 83.33 percent when classifying damage in reinforced concrete structures. Developed to assess surface deterioration from field-acquired color images, the system incorporates an intuitive graphical user interface and operates at an average processing speed of 10.83 frames per second during evaluation.
Aged reinforced concrete structures exhibit various surface deterioration phenomena, such as delamination, leak, and rebar exposure. Conventional visual inspection practices used to monitor these conditions can often be unreliable because of human subjectivity and a persistent shortage of specialized personnel available to conduct detailed physical evaluations.
To address the limitations of manual visual inspection, researchers developed a graphical user interface tool powered by a convolutional neural network and verified its performance with field-acquired images. The workflow prepares input data by normalizing color images of 256 pixels each in width and height, followed by the application of data augmentation techniques prior to feature extraction.
During the feature extraction stage, deep visual characteristics are extracted from the normalized images through three convolutional layers and two pooling layers. Following this feature extraction process, the network classifies the structural damage into three distinct target types: delamination, leak, and rebar exposure. This final classification is carried out using a fully connected layer coupled with a softmax output layer.
The model was trained using the AdamW optimization algorithm, which minimizes a categorical cross-entropy loss function. To enhance generalizability and prevent unnecessary training, the researchers implemented an early stopping rule that automatically halted the training process if the validation loss failed to improve for 20 consecutive epochs.
The complete software application was implemented using Python libraries, specifically TensorFlow, Keras, and Tkinter. The authors designed the software with an intuitive graphical user interface that allows non-experts to carry out damage evaluations and conduct model retraining with a single-button operation.
In K-fold cross-validation using a total of 392 real-world images, the convolutional neural network obtained an average validation accuracy of 82.91 percent, with the best-performing model reaching a validation accuracy of 91.03 percent. In a subsequent evaluation on an independent test set of 60 images not used during training, the system achieved an overall classification accuracy of 83.33 percent, operating at an average processing speed of 10.83 frames per second.
Subject of Research: Automated classification of concrete structural damage using convolutional neural networks
Article Title: Image-Based Techniques for Classifying Damage in Concrete Structures
Article References: Heo, J., Hwang, J.-H., Cho, H.-C., & Ju, H. (2025). Image-Based Techniques for Classifying Damage in Concrete Structures. Architectural Research, 27(4), 157–165. https://doi.org/10.5659/AIKAR.2025.27.4.157
Image Credits: AI Generated
DOI: 10.5659/AIKAR.2025.27.4.157
Keywords: reinforced concrete structures, convolutional neural network, structural damage classification, delamination, leak detection, rebar exposure, visual inspection, graphical user interface, AdamW optimization, categorical cross-entropy, TensorFlow, Keras, Tkinter, K-fold cross-validation, data augmentation, image normalization, early stopping, processing speed
News Source: Cassandra Pierce. (November 16, 2026). Neural Network Tool Classifies Concrete Surface Damage With 83.33 Percent Accuracy. Scienmag.



