An Accurate Deep Key-Point Prediction Model With Low-Level Texture Refinement and High-Level Semantic Enhancement for Bolt Vertex Detection in Industrial Machine Systems
The safety status of intelligent monitoring bolts is crucial for ensuring the stability of industrial building and structural systems. Considering the complexity of the industrial production environment, the collected bolt images may have diverse and complex features, which will seriously affect the...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10909128/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The safety status of intelligent monitoring bolts is crucial for ensuring the stability of industrial building and structural systems. Considering the complexity of the industrial production environment, the collected bolt images may have diverse and complex features, which will seriously affect the recognition accuracy of bolt vertices and easily lead to false alarms and missed alarms of bolt looseness. To address these challenges, this paper proposes a bolt key-point detection method guided by masked-based dual-views model with texture refinement and semantic enhancement. Firstly, inspired by multimodal technology, a dual-views deep fusion method is designed based on the original and relief views to fully utilize their respective advantages to obtain robust key features. Secondly, low-level texture refinement and high-level semantic enhancement modules are designed to improve the edge texture and high-level semantic details of the bolt area, respectively. Finally, we established the gradient update of the above network for key-point recognition loss, target detection loss and cross entropy loss. In addition, we introduced a masked-based unsupervised pre-training paradigm based on convolutional structure to enhance the feature representation ability of the above model. In experiments and discussions, we analyzed the effectiveness of the proposed method in bolt vertex recognition tasks and achieved 0.987 AP and 0.743 Acc. In real-world application cases, a large number of experimental results have demonstrated the necessity of the proposed method. |
|---|---|
| ISSN: | 2169-3536 |