Multiresolution Gray Scale and Rotation Invariant Texture Classification with Local Binary Patterns

Motivation(s)

Analysis of 2D textures have many potential applications (e.g. industrial surface inspection, remote sensing, biomedical imaging). Zernike moments is one of the first studies to include invariance with respect to spatial scale, rotation, and gray scale at the cost of increased computational complexity. Another work focuses on gray scale and rotation invariant texture classification. The gray scale invariance is achieved through histogram equalization of the input image, which cannot correct intra-image (local) gray scale variations.

Proposed Solution(s)

The authors propose a gray scale and rotation invariant texture operator (see Figure 1) based on local binary patterns (LBPs). The LBPs can be thought of as detecting microstructures (e.g. edges, lines, spots, flat areas). The number of histogram bins were chosen such that an average of 10 entries per bin is satisfied. The entries in the bins are the LBP of each pixel.

Evaluation(s)

Training on the Outex image database (using only one lighting angle) achieved a classification accuracy between 80-98%. After a certain amount of granularity, increasing the discretization factor did not significantly improve accuracy.

Future Direction(s)

  • Would extending LBPs to deal with distributions (i.e. spectral distribution) improve classification?

  • How can be histogram be estimated using sparse sampling and still maintain all the desired invariants?

  • What kinds of LBPs would a neural network learn?

Question(s)

  • Does the performance change when the generic chi-square statistic is used instead of the log-likelihood ratio?

  • Multiresolution analysis is easy in this case, but the proposed criterion barely improved accuracy?

Analysis

LBPs are biologically-inspired features that can be computed efficiently and stored as a histogram. The authors recognized that certain LBPs are fundamental properties of local image texture. These LBPs have uniform appearance in the sense that there is a limited number of transitions or discontinuities in the circular presentation of the pattern.

One minor issue with the paper is the poorly worded results section. More illustrative figures and concise descriptions would be appreciated. The underlying theory and and computation seems amenable to GPU acceleration.

It is surprising that the uniformity measure of 2 usually captured 90% of the texture variations. It would have been interesting if the author compared their simple circular pattern against the patterns a human eye exhibits.

Notes

  • Achieving Gray Scale Invariance

    • Invariant against gray scale shifts

      • Subtract the center pixel’s value from each of the surrounding neighbors’ pixel values.

    • Invariant against scaling of gray scale

      • Consider just the signs of the differences i.e. pass the pixel differences through the Heaviside step function.

    • Invariant against any monotonic transformation of the gray scale

      • Assign a power of two for each thresholded sign difference i.e. each result is a different bit location.

  • Achieving Rotation Invariance

    • Naive Solution: Rotate the neighbor set clockwise until the maximal number of the most significant bits are 0.

      1. Occurrence frequencies of individual LBPs vary greatly.

      2. Crude quantization of the angular space at \(45^\circ\) intervals, which can be fixed via increasing the quantization factor \(P\) and radius \(r\).

    • Uniformity Measure

      • The number of 0 to 1 or 1 to 0 transitions.

    • See (9) and (10) for the uniform pattern.

      • The histogram of these outputs accumulated over a texture sample is the final texture feature.

  • Trading Gray Scale Invariance for Local Contrast

    • Divide the previous texture measure by the local variance computed from the pixel values on the circular chain.

References

OPM02

Timo Ojala, Matti Pietikainen, and Topi Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on pattern analysis and machine intelligence, 24(7):971–987, 2002.