An autonomous video analysis method for crack detection on metallic surfaces based on texture recognition and Bayesian data fusion
Journal
Congress on Computing in Civil Engineering, Proceedings
Pages
420-427
Date Issued
2017
Author(s)
Abstract
Aging nuclear power plants are susceptible to the damage onset. Regular inspection of reactor internal components is needed to ensure safe operations. Current practice involves operators watching inspection videos that are captured under the water, and manually labeling the cracks on metallic surfaces of nuclear reactors. This approach is time-consuming, tedious, and subjective. On the other hand, most of the prevalent autonomous crack detection methods were designed for concrete surfaces, which could not detect the cracks on metallic surfaces since they are typically very tiny with low contrast. In addition, the state-of-the-art methods tend to detect scratches, welds, and grind marks on the surfaces as cracks, which cause a large number of false positives. To resolve the above issues, a novel autonomous crack detection method is proposed based on texture analysis and Bayesian data fusion. In each video frame, local binary pattern (LBP) is used to analyze texture and detect crack patches to form crack bounding boxes. After that the crack bounding boxes in different frames are fused with Bayesian decision theory, which dramatically increase the robustness and reliability of detection. To optimize the computation speed, integral histogram and 2-stage support vector machine (SVM) are applied. The proposed method is evaluated using a thousand of frames from several inspection videos. The results showed that it is accurate and efficient, where the other methods could not perform well. ? 2017 American Society of Civil Engineers.
Subjects
Computation theory
Data fusion
Decision theory
Nuclear fuels
Nuclear power plants
Nuclear reactors
Reliability theory
Support vector machines
Surface measurement
Textures
Bayesian data fusions
Bayesian decision theory
Integral histogram
Local binary patterns
Reactor internals
Regular inspections
State-of-the-art methods
Texture recognition
Crack detection
Type
conference paper