Fig. 1. Faster R-CNN deep learning model based on VGG16

Fig. 2. Structure of RPN network

Fig. 3. Clustering. (a) Relationship between k value and accuracy; (b) clustering resultwhen k=12

Fig. 4. Results of four training losses. (a) VGG16 model; (b) resnet50 model; (c) resnet101 model; (d) resnet152 model

Fig. 5. Training loss of different feature extraction network models

Fig. 6. Relationship between model precision and number of iterations. (a) VGG16 model; (b) resnet50 model; (c) resnet101 model; (d) resnet152 model

Fig. 7. Relationship between model recall and number of iterations. (a) VGG16 model; (b) resnet50 model; (c) resnet101 model; (d) resnet152 model

Fig. 8. Target location results of contact network pipe cap. (a) VGG16 positioning results; (b) K-means+VGG16 positioning results; (c) resnet50 positioning results; (d) K-means+resnet50 positioning results; (e) resnet101 positioning results; (f) K-means+resnet101 positioning results; (g) resnet152 positioning results; (h) K-means+resnet152 positioning results

Table 1. Parameters of VGG 16 network convolution process

Table 2. Comparison of deep learning network