Image prediction based on neighbor-embedding methods

Area of Science:

• Image processing
• Data compression
• Machine learning

Background:

• Intra-image prediction is crucial for efficient image coding.
• Existing methods like H.264/AVC intraprediction and template matching have limitations.
• Dimensionality reduction techniques offer potential for improved prediction accuracy.

Purpose of the Study:

• To introduce and evaluate two new intra-image prediction methods.
• To leverage nonnegative matrix factorization (NMF) and locally linear embedding for prediction.
• To analyze the impact of parameters like 'k' and constraints on prediction performance.

Main Methods:

• Developed two intra-image prediction algorithms based on NMF and locally linear embedding.
• Approximated image blocks using linear combinations of k-nearest neighbors.
• Analyzed the influence of the 'k' parameter and nonnegativity/sum-to-one constraints.
• Integrated and evaluated methods within a complete image coding-decoding framework.

Main Results:

• Achieved prediction gains up to 2 dB over H.264/AVC intraprediction.
• Demonstrated gains up to 3 dB compared to template matching.
• Showed gains up to 1 dB relative to a sparse prediction method.
• Evaluated Rate-Distortion (RD) performance of the new methods.

Conclusions:

• The proposed NMF and locally linear embedding-based prediction methods offer significant performance improvements.
• These methods provide a valuable alternative for enhancing image coding efficiency.
• The study highlights the effectiveness of dimensionality reduction in intra-image prediction.