Abstract: This paper presents a concrete crack detection method based on cascaded neural networks in complex environments, aiming at the problems of the traditional deep learning crack detection method in complex environments: low robustness, poor edge area identification accuracy, and large errors in damage quantification results. In this three-step method, first it uses the improved semantic segmentation model to preliminarily identify cracks in complex environments, and determines roughly a cracking area of interest in the image. Then, it optimizes the rough segmentation image using the mask based on pyramid pooling to accurately capture the context information of the crack edge. Finally, it calculates crack width using the image pixel resolution with the QR code targets and a crack width parameter acquisition algorithm. The test results show that compared with the traditional crack identification, this method improves significantly in the five evaluation indicators－precision rate, recall rate, accuracy rate, F1 score and intersection ratio－and achieves an overall detection accuracy of higher than 95%, thereby realizing the detection and quantitative analysis of concrete cracks in complex environments.

Key words: concrete crack detection, complex environment, cascaded neural networks, crack identification, mask optimization