Downstream Semantic Segmentation Model for Low-Level Surface Crack Detection

Figure 3

The precision-recall curves regarding our downstream models built upon DeepLabV3-ResNet101. The best F-measure (ODS F-measure) of each graph is marked with a small rectangle. (a–d) and (e–h) represent two training strategies which are training from the pretrained weights and training from random weights, respectively. Testing results are shown for CrackTree260 in (a, e), CRKWH100 in (b, f), CrackLS315 in (c, g), and Stone331 in (d, h).

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)