Locally rotation, contrast, and scale invariant descriptors for texture analysis

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Real-world image textures exhibit variations in brightness, contrast, scale, and skew due to changing imaging conditions.
• Texture recognition requires similarity measures invariant to these properties, particularly local invariances for undulating surfaces.
• Existing algorithms often involve feature point detection and geometric normalization, with recent focus on local scale and affine invariance.

Purpose of the Study:

• To develop a texture recognition method with local scale and affine invariance.
• To introduce a novel family of linear filters for scale-invariant texture description.
• To enhance texture discrimination and retrieval performance.

Main Methods:

• Utilized invariant combinations of linear filters to create a novel family of scale-invariant filters.
• Developed a texture description invariant to local orientation, contrast, scale, and robust to local skew.
• Employed the A2 similarity measure on histograms derived from filter responses for texture discrimination.

Main Results:

• The proposed filter family enables local scale invariants without scale selection or numerous filters.
• The texture discrimination method outperforms existing approaches on standard datasets (Brodatz, UIUC).
• Demonstrated superior retrieval and classification results, especially for databases requiring local invariance.

Conclusions:

• The novel linear filter-based method provides robust local scale and affine invariance for texture recognition.
• This approach significantly improves texture discrimination accuracy compared to prior methods.
• The technique is effective for real-world image analysis where texture properties vary locally.