Study of the effect of image preprocessing on the performance of a convolutional neural network in classifying concrete crack images

Authors

  • ماهر علي إبراهيم قسم هندسة تكنولوجيا المعلومات-كلية هندسة تكنولوجيا المعلومات والاتصالات-جامعة طرطوس
  • جيانا جمال حسن قسم هندسة تكنولوجيا المعلومات-كلية هندسة تكنولوجيا المعلومات والاتصالات-جامعة طرطوس

Keywords:

Convolutional neural network, CLAH, Bilateral filtering, concrete, Gaussian smoothing, Homomorphic filtering.

Abstract

Convolutional Neural Networks (CNNs) are among the most prominent deep learning models, and one of the most important areas of recent research is the detection of defects and cracks in concrete structures. Given the importance of the initial data preparation phase before network training, this study aims to evaluate the effect of image preprocessing on the performance of a CNN model in classifying concrete images.

A readily available dataset of concrete images from buildings was used to create a convolutional neural network (CNN) model to classify the concrete images into two categories: cracked (C) and uncracked (UC). The model was trained on this data after it underwent several preprocessing steps.

The network was trained for 10 epochs and the images were fed into the convolutional network in four separate inputs so that four models were trained, which will be detailed later. The test accuracy for the four models was (97.10%, 95%, 97.15%, 96.35%) respectively. The results showed that simple processing techniques such as grayscale conversion and Gaussian smoothing performed best because they preserved the original features of the cracks, while complex methods such as Bilateral filtering, CLAH and Homomorphic filtering led to distortion of fine crack details or noise amplification.

Downloads

Published

2026-10-29