Rail Infrastructure Defect Detection Through Video Analytics
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Compared with the traditional railway infrastructure maintenance process, which relies on manual inspection by professional maintenance engineers, inspection through automatic video analytics will significantly improve the working efficiency and eliminate the potential safety concern by reducing physical contact between maintenance engineers and infrastructure facilities. However, the defect does not always have a stable appearance and involves many uncertainties exposed in the clutter environments. On the other hand, various brands of the same devices are used widely on the railway, which shows diverse physical models. Therefore, it creates many challenges to the existing computer vision algorithms for defect detection. In this thesis, two key challenges are abstracted about video/image analytics using computer vision techniques for railway infrastructure defect detection, resulting from the fine-grained defect recognition and the limited labelled learning (few-shot learning). This thesis summarizes the works that have been conducted on utilizing different methods to solve the two challenges.
The first challenge is fine-grained defect recognition. For railway infrastructure defect inspection, damaged or worn equipment defects are usually found in some small parts. That is, the differences between the defective ones and standard ones are fine-grained. Finding these subtle defects is a fine-grained recognition problem. This thesis proposes a bilinear CNNs model to tackle the defect detection problem, which effectively captures the invariant representation of the dataset and learns high-order discriminative features for fine-grained defect recognition. Another challenge is the limited labelled data. In many scenarios, how to obtain abundant labelled samples is laborious. For example, in industrial defect detection, most defects exist only in a few common categories, while most other categories only contain a small portion of defects. Moreover, annotating a large-scale dataset of defects is labour-intensive, which requires high expertise in railway maintenance. Thus, how to obtain an effective model with sparse labelled samples remains an open problem. To address this issue, this thesis proposes a framework to simultaneously reduce the intra-class variance and enlarge the inter-class discrimination for both fine-grained defect recognition and general fine-grained recognition under the few-shot setting. Three models are designed according to this framework, and comprehensive experimental analyses are provided to validate the effectiveness of the models. This thesis further studies the few-shot learning problem by mining the unlabelled information to boost the few-shot learner for defect/general object recognition and proposes a Poisson Transfer Model to maximize the value of the extra unlabelled data through robust classifier construction and self-supervised representation learning.
Please use this identifier to cite or link to this item: