Leveraging Feature Fusion and Convolutional Neural Networks for Concrete Crack Prediction

Detection of cracks in concrete is crucial for the safety of bridges and the overall infrastructure. This paper presents a new hybrid method that combines handcrafted and deep features to significantly improve classification accuracy. Texture and semantic information are captured using Local Binary Patterns (LBP) and a pre-trained Xception model, respectively. These features are converted by the Bag of Visual Words (BoVW) method, combined, and the best features are selected by using the Apriori algorithm. The selected features are classified utilizing the light MobileNetV3, a large network. Our method is tested on four public datasets, namely CODEBRIM, DIMEC, Crack, BCD, and Bridge, using 10, fold cross, validation. The model we introduced has drastically diminished the error rate by less than 1%. Furthermore, it performed well on the accuracy metric, achieving scores of 0.9995, 0.9983, 0.9998, and 0.9993, respectively. Precision values reached 0.9998, 1.0000, 1.0000, and 0.9991, while recall figures stood at 0.9995, 0.9938, 0.9998, and 0.9986 for these same datasets. Compared to other deep learning models trained on the same datasets, our model shows highly encouraging and promising outcomes.