*5.2. Results*

In this subsection, we show the evaluation results and discuss some important details of the proposed model.

#### 5.2.1. Quantitative Analysis

Table 3 shows the detection rate of all defects. From Table 3, we can compare the defect detection performance of our method and comparative methods (CM1-CM7). In these metrics, IoU, which is the standard metric of the semantic segmentation field, is the most important value to evaluate the total performance. We can see that PM obviously outperformed all CMs in this metric.

Next, Table 4 shows the recall rate of detection of each defect. From Table 4, we can observe the specific defect detection performance of our method and comparative methods (CM1-CM7). It should be noted that the metric Recall was used for the evaluation of each defect detection performance since the small crack defects were directly included. For the evaluation of the detection performance of cracks, IoU is not the best evaluation metric because of the difficulty of pixel-level matching. Moreover, considering the application situation, over-detection is considered preferable to miss-detection for the detection of defects. From the above reasons, we selected the evaluation metric Recall in this evaluation.

**Table 3.** Defect detection performance of the proposed method (PM) and the comparative methods (CMs).


**Table 4.** Recall of all kinds of defects in each method.


The proposed method outperforms all comparative methods. According to Table 3 and Table 4, we can further discuss the importance of each component.
