Intelligent Detection Algorithm for Concrete Structure Cracks Based on Deep Learning

Correctly identifying cracks in concrete is essential for their safe use. However, the efficiency of traditional methods of detecting cracks is low. They also suffer from a high degree of subjectivity and therefore make it difficult for existing deep learning algorithms to identify crack locations with a high level of precision (pixel-level accuracy), identify small cracks, and reduce the interference created by complex backgrounds. To address these problems, this research presents a novel form of intelligent identification of cracks in concrete based on U-Net, with the goal of achieving an accurate pixel-level segmentation and quantitative measurement of geometric properties. It employs ResNet-34 as an encoder backbone for multi-scale feature extraction and applies a pyramid clustering model on the encoding side to enable determination of the overall context within the receptive field. Multi-scale average pooling and bilinear interpolation upsampling allow greater ability to distinguish between fine cracks and blocky disruptions. This research also incorporates an attention mechanism from a convolutional neural network into the skip connections by employing two sub-models (channel attention and spatial attention) that adaptively modify the image in order to classify surface texture noise while emphasizing the regions with cracks. A progressive upsampling strategy of the decoder fuses deep semantic information and shallow positional information, and finally outputs the predicted crack probability map through Sigmoid activation. Post-processing techniques such as threshold binarization and morphological opening and closing operations are used to automatically quantize crack length, average width, and maximum width. Results show that the proposed method achieves an mIoU of 82.8%, a recall of 90.9%, and an F1-Score of 89.6%, significantly outperforming the comparative methods. In geometric parameter measurements, the mean relative error for length is 3.2%, and the mean relative error for width is 5.1%, with determination coefficients R² of 0.972 and 0.945, respectively. Even under non-ideal conditions such as strong light overexposure, weak light underexposure, Gaussian noise (σ=25), and tilted shooting (30°), the mIoU remains above 84.7%. The study shows that the solution supplied achieves a good performance trade-off in terms of detection accuracy, robustness and computational effectiveness to provide reliable technical support to enable and maintain the intelligent operation of concrete infrastructure.