Anomaly detection in radiographic images of composite materials via crosshatch regression.

摘要

The development and testing of new composite materials is an important area of research supporting advances in aerospace engineering. Understanding the properties of these materials requires the analysis of material samples to identify damage. Given the significant time and effort required from human experts to analyze computed tomography (CT) scans related to the non-destructive evaluation of carbon fiber materials, it is advantageous to develop an automated system for identifying anomalies in these images. This thesis introduces a regression-based algorithm for identifying anomalies in grayscale images, with a particular focus on its application for the analysis of CT scan images of carbon fiber.;The algorithm centers around a "crosshatch regression" approach in which each two-dimensional image is divided into a series of one-dimensional signals, each representing a single line of pixels. A robust multiple linear regression model is fitted to each signal and outliers are identified. Smoothing and quality control techniques help better define anomaly boundaries and remove noise, and multiple crosshatch regression runs are combined to generate the final result. A ground truth set was created and the algorithm was run against these images for testing. The experimental results support the efficacy of the technique, locating 92% of anomalies with an average recall of 88%, precision of 78%, and root mean square deviation of 11.2 pixels.