TL;DR
This paper compares two methods for measuring structural deformation: Digital Image Correlation (DIC) and Digital Image Processing (DIP) using edge detection. While DIC is generally more accurate, DIP with Canny or Zernike algorithms offers comparable results with significantly faster computation, making it a viable alternative for scenarios with large displacements or limited computational resources.
What Problem Does This Paper Address?
Structural health monitoring is crucial for civil engineering, and accurately measuring displacement is a key aspect. Traditional methods often involve physical sensors, but computer vision-based techniques like Digital Image Correlation (DIC) offer non-contact, high-accuracy measurements. However, DIC can struggle with large displacements or low frame rates, leading to increased computation time, memory usage, and even failure to converge. This paper addresses the limitations of DIC by exploring Digital Image Processing (DIP) techniques, specifically edge detection algorithms (Canny and Zernike moments), as a potentially faster and more robust alternative for structural deformation testing.
Methods
The study compares three algorithms:
- Digital Image Correlation (DIC): A widely used full-field, non-contact method that tracks patterns in images to measure displacement. The paper uses a combination of Fast Normalized Cross-Correlation (FNCC) for whole-pixel search and an iterative approach (FN-AR or FN-GN) with dual cubic interpolation for sub-pixel accuracy.
- Canny Edge Detection: A popular edge detection algorithm that uses Gaussian filtering, gradient computation, non-maximal suppression, and hysteresis thresholding to identify edges. This method is applied to detect and track features (concentric circles) for displacement measurement.
- Zernike Moments: A sub-pixel edge detection method that utilizes orthogonal moments to describe image features. The paper employs Zernike moments to identify sub-pixel edges of concentric circle targets for displacement calculation.
The comparison involves numerical simulations, laboratory experiments using a stepper displacement test system, and a real-world bridge experiment. Performance is evaluated based on accuracy (mean error, RMSE) and computational speed.
Datasets
The paper does not introduce new datasets but utilizes simulated images and real-world experimental data:
- Simulated Images: Generated for numerical simulations, including images with added Gaussian white noise at various levels (1-5%) and translated images of targets (both random speckle patterns for DIC and concentric circles for DIP).
- Laboratory Experimental Data: Acquired using a stepper displacement test system with a target attached to a slider, moved in controlled increments (1m to 5m).
- Real Bridge Experiment Data: Collected from a vehicular bridge, involving video recordings of the bridge span during vehicle passage and simultaneous displacement measurements from a Linear Variable Displacement Transducer (LVDT).
Key Insights and Ideas
- DIP as a Viable Alternative: Edge detection methods (Canny and Zernike) can provide a competitive alternative to DIC for structural deformation monitoring, especially when computational speed is a critical factor.
- Trade-off Between Accuracy and Speed: While DIC generally offers higher accuracy and stability, DIP algorithms are significantly faster, particularly for large displacements where DIC's search range increases computational cost dramatically.
- Noise Sensitivity and Filtering: Both DIP algorithms are sensitive to noise, but Gaussian filtering can improve their performance. DIC shows good noise immunity, though filtering offers less significant improvement.
- CUDA Acceleration for Zernike: Parallelizing the Zernike moment algorithm with CUDA significantly boosts its speed, making it comparable to the Canny algorithm and much faster than DIC.
- Practical Applicability: The study validates the effectiveness of DIP-based methods in real-world scenarios, demonstrating their potential for structural health monitoring.
Key Results
- Accuracy: In simulations, DIC had the smallest mean error (≤0.03 pixel) and RMSE (≈0.2 pixel), while Canny and Zernike algorithms had errors around 0.06 pixel and RMSE around 0.4 pixel. In laboratory tests, DIC showed lower mean error and variance compared to Canny and Zernike. In real bridge tests, DIC exhibited the best accuracy and stability, with maximum errors and RMSEs generally lower than Canny and Zernike, though all methods were within acceptable engineering ranges.
- Speed: For a single point calculation with a search range of 50 pixels, DIC took 3678 ms, Canny took 0.95 ms, and Zernike took 12.91 ms (without acceleration). With CUDA acceleration, Zernike's speed improved dramatically, becoming comparable to Canny.
- Noise Immunity: DIC demonstrated the best noise immunity. Canny and Zernike were more sensitive to noise, but Gaussian filtering improved their performance.
- Large Displacements: DIP algorithms showed advantages in handling large displacements where DIC's computational cost escalates.
Limitations and Future Work
- Accuracy vs. Speed Trade-off: While DIP is faster, it generally exhibits slightly lower accuracy and stability compared to DIC, especially in scenarios with very small displacements or high precision requirements.
- Zernike Moment Complexity: The inherent computational complexity of Zernike moments is noted, though mitigated by CUDA acceleration.
- Future Work: The paper implicitly suggests further optimization of DIP algorithms and their integration into smart hardware for edge computing applications in structural health monitoring. The study provides a foundation for applying non-contact testing systems in specialized scenarios.
Summary updated