Abstract
Automated identification of concrete cracks is essential for ensuring infrastructure integrity and sustained safety. However, the irregularities and non-uniformity of concrete cracks present challenges in achieving a thorough, automated crack assessment meeting current specifications while reducing human involvement. Although relying on crack width to evaluate the degree of cracking has been widely accepted, the measurement of crack width is still limited by the measurement position, and the detection results often require additional processing, lacking intuitive representation. This paper proposes an innovative approach for a complete end-to-end detection system for assessing concrete crack grades, which includes a detection methodology and an evaluation framework. The detection method employs 227×227×3 detection cells trained using convolutional neural networks, each assigned distinct labels to slide and identify cracks. Simultaneously, the evaluation framework translates the concrete current specifications into computer-executable procedures and generates the quantitative evaluation of single and multiple cracks in concrete. The experimental findings indicate the models using fine-tuned AlexNet and VGG16 trained on the self-generated Crack-1000 dataset, generate a low error recognition rate of 5.2% and 2.4% respectively. The applications demonstrate the trained model and evaluation framework can accurately identify crack paths and provide a crack label that matches the actual situation. Compared with traditional methods, the method proposed in this article has superior recognition accuracy and end-to-end detection ability, eliminating the necessity of manual parameter adjustment. This innovative visual solution provides an effective method for detecting and evaluating cracks in concrete structures.