# Bridge Deck Delamination Segmentation Based on Aerial Thermography Through Regularized Grayscale Morphological Reconstruction and Gradient Statistics

## TL;DR

This paper introduces a novel iterative segmentation method using grayscale morphological reconstruction and gradient statistics to accurately detect bridge deck delamination in thermographic images, even with non-uniform backgrounds.

## Contribution

The study presents a new regularized morphological reconstruction framework that effectively handles environmental noise and non-uniform backgrounds in thermographic delamination detection.

## Key findings

- Successfully applied to lab and in-service bridge decks
- Outperforms conventional thresholding and clustering methods
- Handles non-uniform background conditions effectively

## Abstract

Environmental and surface texture-induced temperature variation across the bridge deck is a major source of errors in delamination detection through thermography. This type of external noise poses a significant challenge for conventional quantitative methods such as global thresholding and k-means clustering. An iterative top-down approach is proposed for delamination segmentation based on grayscale morphological reconstruction. A weight-decay function was used to regularize the reconstruction for regional maxima extraction. The mean and coefficient of variation of temperature gradient estimated from delamination boundaries were used for discrimination. The proposed approach was tested on a lab experiment and an in-service bridge deck. The result showed the ability of the framework to handle the non-uniform background situation which often occurred in practice and thus eliminates the need for inferencing the background required by existing methods. The gradient statistics of the delamination boundary in the thermal image were indicated as the valid criterion refine the segmentation under the proposed framework. Thus, improved performance was achieved compared to conventional methods. The parameter selection and the limitation of this approach were also discussed.

---
Source: https://tomesphere.com/paper/1906.03723