In the present work, the potential of EGs as a crack pattern descriptor is illustrated by means of a simple yet comprehensive analysis process, specifically developed for demonstration purposes. This process aims at the analysis of crack images captured during the tensile strength test of reinforced cementitious renders used as a base coat for building thermal insulation systems. The typical procedure for this laboratory test requires the manual identification and evaluation of cracks in a sequential image series acquired from a sample under increasing unidirectional tensile strain. This is a particularly challenging task for automated image analysis, because thin cracks must be identified and measured on a rough surface, which causes a significant texture in the resulting image, imposing severe requirements both on the image acquisition system and on the analysis process. Moreover, the cracks identified in the whole image sequence set must be correlated in order to reconstruct the crack evolution during the test. In the present work, the crack detection is implemented with a simple threshold-based binary segmentation process. The resulting binary image is then skeletonized in order to derive a raw, primal EG that is further processed to provide a refined, final EG. The advantages of the EG-based crack description are apparent even in the early processing of the primal EG and are highlighted by illustrating an implementation that includes noise deletion, subgraph merging and artefact removal. The reported process, albeit a relatively simple example, allows the generation of a fairly good final EG that describes the crack pattern in the image and constitutes the basis for the following analyses. The quality of the obtained final EG is assessed by using an EG-based reference ground truth with traditional metrics specifically adapted to the comparison of EGs. Examples of crack analyses carried out directly on the final EG (crack length and crack branching) and on the original grayscale image by direction of crack EG (crack width) are described alongside an image-to-image crack correlation analysis in the whole time-series image set. The advantages obtained in the refinement of the crack paths and in the following crack analyses are discussed and compared with the traditional implementation based on the direct analysis of the binary segmented image.
Graph 35 Manager Plus Crack
DOWNLOAD: https://tinurll.com/2vH0t2
It is worth noting that the EG representation of a crack network can be handled with a wide variety of methods thanks to its compact description of both topological and geometrical crack skeleton features, allowing the development of a broad family of procedures for accurate crack pattern refining and analysis. Some of these are described in the present work, but this is only a small set of possible processing strategies made available by this approach, which allows to merge the large collection of algorithms developed for graph processing with the extensive variety of digital image analysis methods reported in the literature.
Complete final EG generation from a 1.5% strained sample image. Original image (a); binary segmentation (b); primal EG after skeletonization (c); noise removal (d); subgraph merging, recovery, filtration and removal of crack artefacts (e); final EG after crack correlation (f). Colors indicate different subgraphs/cracks. Distal edges not shown.
In the sample depicted in Figure 4, the described noise removal process results in the deletion of >109 103 vertices (>69%) and >58 103 subgraphs (>99.3%), leading to a near complete noise elimination but substantially retaining almost all crack features (Figure 4d). After noise removal, the EG comprises 47.8 103 vertices grouped in 359 subgraphs.
An example of incomplete capture of crack pattern is reported in Figure 4d, where the relatively clean representation of the crack paths demonstrates a deep fragmentation of each crack into several subgraphs (depicted with different colors) because of the small gaps formed during the binarization process. The above-described process provides a quite remarkable result, decreasing the count of subgraph from 359 to 25 and collating most of the visible cracks.
A particularly severe example of subgraph loss is depicted in Figure 8, where a small crack image resulted in a very dispersed collection of small subgraphs (Figure 8b), with large parts canceled afterwards by the noise removal algorithm and thus substituted by distal edges by the subgraph merging process (depicted in gray in Figure 8c). The subgraph recovery process finds the legitimate small crack subgraphs in the primal EG and inserts them into the merged subgraph, with near complete reconstruction of the crack recorded in the original image (Figure 8d). It is, however, important to note that this process is aimed at the recovery of deleted subgraphs in the internal part of a crack path and it cannot retrieve missing small subgraphs located in the crack ends (Section 3.3.7).
Small subgraph recovery. Image of a small crack (a) generates a primal EG (b) that is processed up to the subgraph merging step (c), where several parts were filtered out during noise removal and replaced by distal edges. The recovering algorithm retrieves the missing small subgraphs from primal EG and inserts them in the merged subgraph (d). Distal edges are depicted in gray.
In the sample depicted in Figure 4, the subgraph recovery step significantly improved the crack EG completeness, recovering >2300 vertices that add to the >47 103 vertices (+5.0%) previously present in the EG.
In the reported sample, the processing of the crack EG with the described algorithm allows to identify two cases (corresponding to the four upper cracks in Figure 4e) and, after a subgraph joining operation, the resulting final EG is composed of 15 subgraphs, each correlated with a single crack. Moreover, during the joining process, a further search for the recovery of small subgraphs (as described in Section 3.3.3 for distal edges) is carried out, resulting in 191 added vertices.
Computed crack genealogy trees for selected cracks in the sample shown in Figure 4 and Figure 15. Each vertex represents a specific crack (sub-graph) in the final EG of the corresponding image. Two crack coalescence events are observed at 0.8% and 1.5% strain. Crack labels as depicted in related figures.
In this work, a feature extraction branch based on graph neural network is added to a typically semantic segmentation network to form a new end-to-end network structure. And experiments are conducted on concrete pavement crack segmentation to evaluate the performance improvement. In this regard, the main contributions of this work can be summarized as follows:(i)A semantic segmentation network framework with graph neural network branch is proposed to segment the concrete pavement crack. The performance of crack segmentation is significantly improved based on the original segmentation network. In addition, the inclusion of the graph branch improves the continuity of crack segmentation.(ii)A generation method to convert images into graphs is designed, which enriches the feature map dimension of images.(iii)A new dataset of 3D concrete pavement crack images is established and applied to evaluate the proposed network.
The rest of this paper is organized as follows. Section 2 describes the related research on pavement crack detection and the development of graph neural networks. Section 3 introduces the detailed architecture of the segmentation network with graph neural network branch. Section 4 represents the experiment setting. Section 5 discusses the experiment results. Finally, Section 6 concludes the work and presents the findings of this research.
In this section, the graph neural network feature extraction branch and the main body of semantic segmentation are first introduced, respectively. Then, the proposed network structure for crack detection on the concrete pavement is described.
Figure 6 shows the result of U-Net and GA-Unet at different epochs. The effort of concrete pavement crack segmentation is improving and the results become closer to the ground truth with the epoch increasing regardless of the U-Net or GA-Unet. However, the GA-Unet is more accurate than U-Net for the same training epoch. The addition of the graph branch can improve the learning ability, enhance feature extraction capability, and boost the convergence speed.
The comparison experiment between the AutoEncoder, PSPNet [37], U-Net [19], KiUnet [38], and GA-Unet is conducted, and the results are illustrated in Table 1 and Figure 7. AutoEncoder is the simplest segmentation network with only an encoder and decoder structure. U-Net is the segmentation backbone in KiU-Net and GA-Unet. KiU-Net adds an over-complete representation branch based on U-Net to promote the performance. GA-Unet adds the graph network branch to enrich the feature represents. The U-Net can be regarded as the original semantic segmentation network compared to the GA-Unet. The comparison result between U-Nnet and GA-Unet can verify the validity of graph network branch. The performance is represented by four metrics, and the optimal results have been highlighted in bold in Table 1. GA-Unet achieves the optimal results in the metrics of F1, and IoU, which are 0.53 and 0.37, respectively. In addition, GA-Unet has a significant improvement in Recall, F1, and IoU metrics compare to the U-Net, which is increased by 0.12, 0.08 0.06. Although GA-Unet is weaker than U-Net in terms of Precision and KiU-Net in terms of Recall, GA-Unet achieves better performance in segmenting cracks in concrete pavement in general. Figure 7 shows the comparison between the segmentation image of PSPNet, U-Net, and GA-Unet. The quality of the crack segmentation conducted by GA-Unet achieved better results than U-Net under different conditions. 2ff7e9595c
Comments