Analysis of gear surface morphology based on gray level co-occurrence matrix and fractal dimension

Comparison of three image-analysis-based visual texture calculation methods: energy, entropy, and texture change index

BACKGROUND

Three image analysis methods to measure visual texture were applied to an image with much texture (scaled carp), and one with little texture (mirror carp). For each method, the effect of image rotation at 0°, 10°, 20°, 30°, 45°, 60°, 75° and 90° was evaluated.

RESULTS

Rotation did not change energy (E) and entropy (H) calculations using image histograms. Using co-occurrence matrices with different step size d (1-19 in increases of 2) and step angle θ (0°, 45°, 90° and 135°) showed that, as d increased, E decreased and H increased, and the number of legitimate pixel pairs decreased linearly. Averaging E and H at different θ values rendered the results rotation invariant. Theoretically, the 'texture primitives' method is not rotation independent. However, the variation in texture change index (TCI) with image rotation was negligible. Also, the increase in TCI between the less textured image and the more textured image was 5.3-11. In comparison, the E values from histograms for the images above were 0.0069-0.0081. For co-occurrence matrix-based calculations, at d = 1 and for all θ, E range was from 220 to 389 for scaled carp and from 232 to 412 for mirror carp.

CONCLUSION

The more than doubling of TCI for these images implies that it is a more precise method than energy and entropy to discern between visual texture levels. © 2022 Society of Chemical Industry.