Affine invariant texture analysis based on structural properties

Jianguo Zhang, Tieniu Tan2002

TL;DR

This paper introduces a novel method for texture analysis that is invariant to affine transformations (like rotation, scaling, and skewing). By deriving new structural properties based on texel areas and creating an "area-ratio map," they achieve robust texture classification, even under significant image distortions.

What Problem Does This Paper Address?

In computer vision, accurately analyzing textures is crucial for tasks like image recognition and segmentation. A major challenge is that images are rarely captured from a single, perfect viewpoint. Real-world images often undergo various geometric distortions, such as rotation, scaling, translation, and skewing (collectively known as affine transformations). Existing texture analysis methods often struggle with these distortions, leading to misclassification. This paper aims to develop a texture analysis technique that remains effective despite these affine transformations, making it more practical for real-world applications.

Methods

The core of this paper lies in developing "affine invariant structural properties." The authors propose that the ratio of areas of two triangles (or more generally, two closed planar shapes) remains constant even when subjected to an affine transformation. This is a key theoretical insight.

To apply this to texture analysis, they introduce an "area-ratio map." This map is constructed by first segmenting an image into uniform regions (texels) and then calculating the ratio of the areas of every pair of texels. This creates an $n \times n$ matrix where $n$ is the number of texels.

From this area-ratio matrix, an area-ratio histogram is generated. This histogram represents the distribution of these area ratios and serves as the texture feature. The authors argue that these histograms are invariant to affine transforms because the underlying area ratios are invariant. For classification, they use a k-nearest neighbor algorithm based on these histograms.

Datasets

The paper uses the Brodatz texture album for evaluation. Specific textures mentioned include D005, D012, D020, D051, D056, D074, D075, and D104. They create multiple affine-transformed versions of these textures for training and testing.

Key Insights and Ideas

  • Area Ratios as Affine Invariants: The fundamental insight is that the ratio of areas of geometric shapes (like triangles) is preserved under affine transformations. This provides a mathematical basis for creating invariant features.
  • Area-Ratio Map for Texture Description: Instead of relying on direct correspondence between texels in original and transformed images (which is difficult), the area-ratio map captures the relationships between texels' areas, making it robust to global transformations.
  • Histogram as a Compact Feature: Summarizing these area ratios into a histogram provides a compact and effective feature representation for texture classification.
  • Robustness to Noise: The method demonstrates good performance even when textures are corrupted by significant levels of Gaussian white noise, Salt&Pepper noise, and Speckle noise, indicating practical utility.

Key Results

  • The proposed method achieves high classification accuracy on the Brodatz textures, with an average accuracy of 96.25%.
  • Six out of eight texture classes achieved 100% classification accuracy.
  • The method shows strong similarity between different affine transformations of the same texture (low distance D) and low similarity between different textures.
  • The algorithm exhibits good noise robustness, maintaining around 80% accuracy even at a signal-to-noise ratio (SNR) of 5.

Limitations and Future Work

  • Micro-textures: The method is less suitable for "micro textures" where texels are not easily identifiable or defined. The paper notes that small texels in textures like D005 can be problematic due to quantization errors affecting area ratio calculations.
  • Texel Extraction: While not the primary focus, the performance of the algorithm is dependent on the quality of texel extraction. More sophisticated segmentation techniques could potentially improve results.
  • Future Work: The authors suggest that the detailed texture information within each texel surface could be further exploited, and that the sensitivity to small texels should be addressed in future research to improve accuracy and robustness.

Summary updated

Abstract

This paper presents a new texture analysis method based on structural properties. The texture features extracted using this algorithm are invariant to affine transform (including rotation, translation, scaling, and skewing). Affine invariant structural properties are derived based on texel areas. An area-ratio map utilizing these properties is introduced to characterize texture images. Histogram based on this map is constructed for texture classification. Efficiency of this algorithm for affine invariant texture classification is demonstrated using Brodatz textures. 1.
BibTeX
@misc{zhang2002affine,
  author = {Jianguo Zhang and Tieniu Tan},
  title = {{Affine invariant texture analysis based on structural properties}},
  year = {2002},
}

Loading related content...

Figure