Concrete Crack Identification Framework Using Optimized Unet and I–V Fusion Algorithm for Infrastructure

Abstract

Currently, most of the concrete crack detection models proposed mainly rely on a single deep learning method, whose performance is limited. To solve the problem, this work presents a deep learning framework for crack identification of concrete. First, a histogram equalization method is adopted to processed the original image, which can effectively enhance the contrast and brightness. Then, to extract effective features of the crack, multiple filters are employed for crack detection, which fusion with original data. In addition, the Unet network is employed as the base classifier for initial diagnosis of concrete crack. To raise the extraction precision, enhanced attention mechanism module is applied to the Unet model for parameter optimization. The combination of Dice function and cross-entropy loss function is applied to evaluate the model performance. The voting integration algorithm is utilized to each prediction result for the decision of the final prediction result. Finally, to demonstrate the effectiveness of the proposed method, a total of 608 steel fiber concrete crack images are collected from laboratory. The results indicate that the proposed deep learning framework offers the optimal comprehensive recognition performance.