LP²DH: A Locality-Preserving Pixel-Difference Hashing Framework for Dynamic Texture Recognition

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Spatiotemporal Local Binary Pattern (STLBP) is a common dynamic texture descriptor but suffers from high dimensionality.
• Existing methods often sacrifice inter-plane correlation by extracting features on orthogonal planes.

Purpose of the Study:

• To propose a novel framework, Locality-Preserving Pixel-Difference Hashing (LP2DH), for dynamic texture recognition.
• To address the high dimensionality issue of STLBP while preserving feature discriminative power and local structure.

Main Methods:

• LP2DH jointly encodes pixel differences in the spatiotemporal neighborhood, transforming Pixel-Difference Vectors (PDVs) into compact binary codes.
• It incorporates locality-preserving embedding and a curvilinear search strategy for optimizing hashing.
• Dictionary learning and histogram representation are used for final feature extraction.

Main Results:

• LP2DH achieved state-of-the-art performance on three major dynamic texture recognition benchmarks.
• Achieved 99.80% accuracy on UCLA, outperforming DT-GoogleNet (98.93%).
• Achieved 98.52% on Dyn-Tex++ and 96.19% on YUPENN, surpassing existing methods.

Conclusions:

• The proposed LP2DH framework effectively handles high dimensionality in dynamic texture recognition.
• LP2DH demonstrates superior performance and preserves local structure, offering a robust solution for dynamic texture analysis.