Abstract
Automating outdoor defect detection, particularly for identifying hail damage on roofs, is significantly hindered by surface texture noise and drone-induced motion blur. This paper introduces a novel Bayesian edge detection framework that utilizes "reflected non-local priors" to distinguish true defects from background noise. By framing edge detection as a hypothesis testing problem, the proposed method effectively suppresses false positives that typically plague frequentist approaches like the Canny detector. Experimental results on both simulated and real-world drone imagery demonstrate that the Bayesian detector provides superior noise suppression and aligns more closely with human expert assessments.
Introduction
Automating the inspection of outdoor structures using drones offers significant efficiency gains for the insurance and construction industries. However, two primary obstacles remain: motion blur from the drone's flight and the inherently rough surface textures of materials like asphalt shingles. Standard edge detection tools, such as the Canny detector, rely on local intensity gradients. On textured surfaces, these tools often mistake the "noise" of surface granules for actual defects, leading to a high false discovery rate and unnecessary repair costs.
Methodology: Reflected Non-Local Priors
To address these challenges, the authors move away from gradient-based methods toward a Bayesian hypothesis testing framework. For each pixel, the system tests a null hypothesis ($H_0$: no edge) against an alternative hypothesis ($H_1$: an edge exists).
The core innovation is the introduction of the Reflected Non-Local Prior. Non-local priors are probability distributions that assign zero density to the null hypothesis. The authors created a "reflected" version by subtracting a standard Gaussian density from its peak, effectively creating a "trough" or "hole" at zero. This mathematical structure forces the detector to require substantial evidence before rejecting the null hypothesis, naturally filtering out low-level texture noise and motion blur.
By calculating the Bayes Factor—the ratio of the likelihood of the data under the alternative hypothesis versus the null—the system identifies pixels that are truly part of a defect. This approach allows the model to be "stubborn" against background noise while remaining sensitive to significant structural anomalies.
Key Results and Discussion
The method was evaluated using simulated noisy images and high-resolution drone imagery of asphalt and slate roofs. The Bayesian detector demonstrated a marked improvement in precision ($R_s$) compared to the Canny detector, which produced a dense "cloud" of false edges in high-noise scenarios.
As shown in the figure above, the Bayesian detector maintains a clean outline of surface features and defects, whereas the Canny detector is overwhelmed by the granular texture of the shingles. While the Bayesian method may sacrifice a small amount of recall ($R_t$), it maintains over 80% recovery of true edges, which is sufficient for identifying damage patterns. Furthermore, the detector's output aligned closely with the decisions made by human insurance adjusters, suggesting high practical utility for automated claims processing.
Conclusion and Future Work
The Bayesian approach with reflected non-local priors provides a robust solution for defect detection in noisy, outdoor environments. While the current implementation requires manual tuning of hyperparameters (such as the smoothing factor $\sigma_s$ and the prior shape $\beta$), the detector remains stable across various settings. Future research will focus on integrating machine learning techniques to automatically calibrate these parameters based on the specific texture of the surface under inspection, further enhancing the autonomy of drone-based inspection systems.
Summary updated