Performance evaluation of an improved deep CNN-based concrete crack detection algorithm

Abstract

This study uses a novel directional lighting approach to produce a computationally efficient five-channel Visual Geometry Group-16 (VGG-16) convolutional neural network (CNN) model for concrete crack detection and classification in low-light environments. The first convolutional layer of the proposed model copies the weights for the first three channels from the pre-trained model. In contrast, the additional two channels are set to the average of the existing weights along the channels. The model employs transfer learning and fine-tuning approaches to enhance accuracy and efficiency. It utilizes variations in patterned lighting to produce five channels. Each channel represents a grayscale version of the images captured using directed lighting in the right, below, left, above, and diffused directions, respectively. The model is evaluated on concrete crack samples with crack widths ranging from 0.07 mm to 0.3 mm. The modified five-channel VGG-16 model outperformed the traditional three-channel model, showing improvements ranging from 6.5 to 11.7 percent in true positive rate, false positive rate, precision, F1 score, accuracy, and Matthew's correlation coefficient. These performance improvements are achieved with no significant change in evaluation time. This study provides useful information for constructing custom CNN models for civil engineering problems. Furthermore, it introduces a novel technique to identify cracks in concrete buildings using directed illumination in low-light conditions.

Threshold-Based BRISQUE-Assisted Deep Learning for Enhancing Crack Detection in Concrete Structures

“…Automatic defect detection can be separated into two main approaches: white-box and black-box techniques [11]. The former uses algorithms such as edge detectors and thresholding [12], whereas the latter employs machine learning and artificial neural networks [13].…”