Local energy pattern for texture classification using self-adaptive quantization thresholds.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Texture classification is crucial for image analysis.
• Existing statistical methods have limitations in describing local structures and handling imaging variations.

Purpose of the Study:

• To propose a novel local energy pattern (LEP) method for robust texture classification.
• To evaluate the performance of LEP against established statistical texture descriptors.
• To extend the LEP representation for dynamic texture recognition.

Main Methods:

• Generating local feature vectors using normalized local-oriented energies.
• Quantizing feature vectors with self-adaptive thresholds and N-nary coding.
• Utilizing frequency histograms as the final representation feature.
• Benchmarking on KTH-TIPS and KTH-TIPS2-a material categorization datasets.

Main Results:

• The proposed LEP method demonstrated superior performance on the KTH-TIPS2-a database.
• Competitive results were achieved on the KTH-TIPS database compared to methods like LBP and Weber local descriptor.
• Favorable recognition results were obtained on the UCLA dynamic texture database.

Conclusions:

• The local energy pattern method offers a robust and effective approach for static and dynamic texture classification.
• LEP exhibits improved descriptive power for local image structures and resilience to imaging conditions.
• The method provides a competitive alternative to existing statistical texture analysis techniques.