Automatic Recognition of Concrete Cracks in Industrial and Civil Buildings Based on CNN + SVM/Random Forest

Abstract

Concrete cracks are a key indicator for evaluating the safety of industrial and building structures Civil buildings. Traditional manual detection methods have low efficiency, strong subjectivity, and high labor costs, which cannot meet the needs of large-scale automated monitoring Significant progress has been made in learning based crack detection algorithms, with most existing models For the development of transportation infrastructure (bridges, pavements, tunnels), due to significant differences, the performance will decrease by 15% to 25% when directly migrated to building scenarios In terms of lighting, background texture, and crack morphology. In addition, currently over 95% of research Only focusing on binary classification, unable to provide crack width and Severity is crucial for engineering practice. Although pixel level segmentation methods can Implementing crack quantification relies on extremely expensive manual labeling (every 10-20 minutes) Images make large-scale applications impractical. To address these challenges, this article proposes an automatic crack recognition scheme for lightweight concrete used in industrial and civil buildings, combined with CNN feature extraction based on SVM/random forest classifier. This scheme utilizes CNN Superior generalization of hierarchical feature extraction and SVM/random forest in small samples Plot. The four pre trained CNN backbone networks (VGG16, ResNet50, MobileNetV3, EfficientNet Lite) are System comparison and introduction of ZCA whitening and anti distillation to enhance Has robustness to complex interferences such as shadows and stains. For edge deployment (Smartphones, drones, Raspberry Pi), full frame lightweighting, and model compression Implemented support vector pruning and parameter quantization. It is crucial to have coarse particles By mining the severity of cracks, innovative classification of crack severity was achieved without pixel level segmentation The implicit mapping between CNN features and crack width greatly reduces annotation costs. Experiments on public datasets (SDNET2018, Crack500) and self built industrial/civilian datasets Constructing a crack dataset shows that this approach increased F1 scores by 6% in complex situations Interference scene. The optimized MobileNetV3+compressed SVM model only achieves 92 KB, single image inference time of 0.8 ms on Raspberry Pi 4B, crack severity classification The accuracy rate is 87%. These results fully meet the requirements of practical engineering, Efficient and deployable solutions for building health assessment.