A deep transfer learning-based algorithm for concrete surface defect detection

Abstract The integrity and safety of concrete structures are crucial in modern civil engineering. Traditional image-defect detection methods are significantly affected by light and noise. Methods based on deep learning, particularly convolutional neural networks (CNNs), offer new solutions for automated defect detection. This paper introduces a concrete surface defect detection algorithm based on deep transfer learning. Our model utilizes the ResNet50 architecture as its backbone and incorporates Atrous Spatial Pyramid Pooling (ASPP) to adapt to the diversity of defects and enhance recognition accuracy. In addition, the use of a self-attention mechanism effectively enhances the model’s focus on and analysis of key defect areas, improving its ability to perceive and select features for defects of various sizes, aiding precise localization and classification under complex conditions. Image preprocessing techniques such as resizing, conversion to grayscale, histogram equalization, and noise addition further enhance the robustness of the model in diverse real-world scenarios. On the publicly available COncrete DEfect BRidge IMage Dataset, our model achieved a mean average precision (mAP@0.5) of 0.90, showing a 3.4% improvement over the current best baseline method. Through transfer learning, it reached 0.92 on our custom dataset, significantly outperforming existing methods. These results demonstrate the effectiveness and superiority of the proposed model for detecting concrete defects. Additionally, we evaluated the impact of different architectures on model performance. Compared to traditional models using the VGG16 as the backbone, the ResNet50 architecture used in this study reduces parameter count by 5.41 times while improving performance by 6.1%.