Comparison of texture analysis schemes under nonideal conditions

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Advancements in texture analysis require evaluation in realistic, non-ideal conditions.
• Real-world image acquisition involves variations in illumination and geometric transformations.

Purpose of the Study:

• To assess the performance of texture analysis algorithms under non-ideal image acquisition environments.
• To identify conditions under which novel texture analysis advancements fail or succeed.

Main Methods:

• Evaluated nine popular texture analysis algorithms.
• Used three datasets with varying difficulty levels.
• Conducted experiments under five different non-ideal setups, including illumination variations and affine/non-affine transformations.

Main Results:

• Most state-of-the-art techniques showed poor performance in non-ideal environments.
• Algorithm performance critically depended on the nature of the textural surface.
• Multiscale features were robust to illumination/rotation but failed with scale changes.
• Structured/periodic textures demonstrated stable performance despite variations.

Conclusions:

• Current texture analysis methods are not robust to common real-world image variations.
• Texture surface characteristics significantly influence algorithm reliability.
• Structured textures offer a more stable domain for texture analysis under challenging conditions.