Weakly Supervised Approach to Defect Detection in Concrete Structures Using Hammering Test

Acoustic Inspection of Concrete Structures Using Active Weak Supervision and Visual Information

“…This is certainly due to the higher number of clusters since the two blocks were from different batches and defect types also vary. Without visual information, while the consistency of output was slightly increased by our proposed active query scheme compared to Louhi Kasahara et al [14], the average performance slightly decreased. However, since the addition of visual information allows the proposed method to outperform the method of Louhi Kasahara et al [15], it can be strongly suspected that the nature of errors was different when using the proposed active query scheme: those errors were easier to compensate for by the spatial estimator.…”

“…The methods of Louhi Kasahara et al [14] and Louhi Kasahara et al [15] were weakly supervised. Louhi Kasahara et al [14] only used weak supervision, as opposed to Louhi Kasahara et al [15] and the proposed method that use position information to complement weak supervision. The method of Louhi Kasahara et al [15] assumed that the human user would randomly select pairs of samples for weak supervision.…”

“…Figure 8 illustrates the defect detection outputs using the proposed method. Figure 9 reports the performance obtained using the method of Louhi Kasahara et al [14], RCA with the proposed active query method, the method of Louhi Kasahara et al [15], and the proposed method. In Figure 8a,b, it can be seen that the separation defect/non-defect hammering samples by our proposed method is excellent, only missing a couple of samples.…”

“…The defect/non-defect borders were however well discerned and the degraded surface condition of Area 1 did not seem to negatively affect the performance of our proposed method. RCA with the proposed active query scheme should be put in direct comparison with Louhi Kasahara et al [14], the only difference between those two methods being the active query scheme. For Cases 1 and 2, results in Figure 9a,b show that the proposed active query scheme achieves consistent, higher quality of weak supervision, resulting in a better performance in average and lower values of standard deviation.…”

“…From the point of view of concrete structure inspection, such approaches are also extremely attractive since human involvement can be maintained during the actual inspection process. The work in [14] proposed an initial framework for weakly supervised inspection of concrete structures, later expanded by the addition of visual information in [15]. However, weak supervision was assumed to be provided by the human user on randomly chosen pairs of samples.…”