Context-aware and locality-constrained coding for image categorization.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Bag-of-Features (BOF) based feature design is crucial for image categorization.
• Existing coding strategies for BOF suffer from ambiguity, hindering performance.
• There is a need for more discriminative and robust feature coding methods.

Purpose of the Study:

• To introduce a novel context-aware and locality-constrained Coding (CALC) approach.
• To address the ambiguity in BOF coding procedures for improved image categorization.
• To enhance object description by incorporating contextual information.

Main Methods:

• Learning a word-to-word cooccurrence matrix to capture spatial distribution of local features.
• Measuring context distance between local features and code words using the cooccurrence matrix.
• Developing a coding strategy that considers both feature space locality and context space, with feature weighting.

Main Results:

• The CALC approach semantically preserves information during coding.
• It effectively alleviates noise distortion within each class.
• Experiments on Scene-15, Caltech101, and Caltech256 datasets show significant performance improvements over baselines and state-of-the-art methods.

Conclusions:

• The proposed CALC method offers a superior coding strategy for BOF-based image categorization.
• Incorporating context and locality constraints leads to more discriminative feature representations.
• CALC achieves competitive and often superior results compared to existing advanced techniques.