Introduction
Automatic visual inspection is a cornerstone of modern manufacturing, yet it faces significant hurdles when dealing with "random" textures. While detecting a hole in a uniform fabric is straightforward, identifying a hairline crack in a slab of granite is notoriously difficult. This difficulty stems from the tiny scale of cracks—often occupying less than 1% of the surface area—and the natural variations in random textures that can easily be mistaken for defects.
Researchers from the University of Genoa have proposed a solution: Conditional Texture Anisotropy (CTA). This metric identifies cracks by detecting where a texture’s directional behavior deviates from its expected "normal" state.
Methodology
The CTA approach moves beyond global inspection by focusing on localized directional changes. The pipeline consists of several key stages:
- Feature Extraction: Textural features are calculated from a $16 \times 16$ pixel window using the Fourier Power Spectrum across four spatial frequency bands and four orientations ($0^\circ, 45^\circ, 90^\circ, 135^\circ$).
- Conditioning: Unlike standard anisotropy, CTA is "conditioned" on the expected texture of defect-free material. The system learns the probability distribution of features for each direction from a "clean" sample.
- CTA Calculation: The CTA value compares the maximum and minimum conditional probabilities across orientations. A crack causes the probability of "normal" texture in its specific direction to drop to near zero, creating a sharp spike in the CTA value.
- Classification and Thresholding: Pixels are classified using Mahalanobis distance. A two-threshold approach (similar to hysteresis thresholding) is used to ensure spatial connectivity and reduce noise, labeling "uncertain" pixels as cracks only if they connect to high-certainty crack pixels.
Results and Discussion
The effectiveness of the CTA metric is most evident in its signal-to-noise ratio. When profiling the CTA values across a granite image, the metric remains low (0.2–0.4) for normal texture but spikes dramatically to nearly 1.0 at the exact location of a crack.
This high contrast makes thresholding extremely reliable. The core insight is that a crack acts as a "texture destroyer," disrupting local symmetry in a way that is invisible to global filters but obvious when directional consistency is measured.
Limitations
While robust, the current implementation relies on a Gaussian assumption for feature distributions, which may not hold for all irregular materials. Additionally, the computational intensity of Fourier transforms and Mahalanobis distances poses challenges for real-time industrial applications. Future work may also need to increase the number of orientations to detect very thin cracks that fall between the current $45^\circ$ increments.
Conclusion
By leveraging the inherent directionality of cracks, the CTA method provides a powerful tool for inspecting complex, randomly textured surfaces. It successfully isolates subtle defects that traditional global methods often miss, offering a significant advancement for industrial quality control.
Summary updated