A deep region-based pyramid neural network for automatic detection and multi-classification of various surface defects of aluminum alloys
Aluminum alloys have a wide range of applications in building and civil infrastructure. During the process of production, transportation and storage, various defects inevitably occur on the material, including blisters, scratches, base exposure, dirty points, etc. The efficiency and accuracy of defe...
Saved in:
| Main Authors: | , , |
|---|---|
| 其他作者: | |
| 格式: | Article |
| 語言: | English |
| 出版: |
2022
|
| 主題: | |
| 在線閱讀: | https://hdl.handle.net/10356/159837 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 機構: | Nanyang Technological University |
| 語言: | English |
| 總結: | Aluminum alloys have a wide range of applications in building and civil infrastructure. During the process of production, transportation and storage, various defects inevitably occur on the material, including blisters, scratches, base exposure, dirty points, etc. The efficiency and accuracy of defect detection and classification can be greatly improved by replacing the conventional manual approaches with modern deep learning techniques. This paper proposes to use computer vision and deep learning techniques to achieve automatic detection of various defects of aluminum alloys. Faster region-based convolutional neural network (Faster R–CNN) is selected as the fundamental framework due to its advantages in efficiency and accuracy. According to the characteristics of defects in aluminum alloys, the framework is optimized by (1) feature pyramid networks (FPN) for integration of low-level structural information with high-level semantic information, as well as increasing the feature mapping resolution of small targets; (2) deformable-ConvNets for feature extraction at the most appropriate places; and (3) contextual ROI pooling for fine adjustment of region proposal taking the entire image as a reference. To make full use of the limited samples, the training process is also optimized by (1) utilizing samples without defects; and (2) sample duplication by horizontal and vertical rotation. The proposed approach is validated on a dataset with 10000 images and is shown to have outstanding performance compared to other existing deep learning approaches in defect detection and classification. |
|---|