Local coding based matching kernel method for image classification.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing visual similarity metrics face a trade-off between efficiency (Bag of Visual Words - BoV) and effectiveness (kernel-based methods).
• Euclidean distance-based local kernels are often suboptimal for local features like SIFT and HoG, which exhibit heavy-tailed distributions.
• The need for scalable and efficient visual matching in large image datasets is critical.

Purpose of the Study:

• To develop a unified framework for visual matching that integrates both BoV and kernel-based approaches.
• To propose a novel, efficient local coding based matching kernel (LCMK) method.
• To address the limitations of Euclidean distance metrics for heavy-tailed feature distributions.

Main Methods:

• A unified visual matching framework is proposed, incorporating local kernels.
• A novel Local Coding based Matching Kernel (LCMK) method is introduced, leveraging manifold structures in Hilbert space.
• The LCMK method exploits recent advances in feature coding techniques for improved performance.

Main Results:

• The proposed LCMK method achieves linear computational complexity, enabling efficient and scalable visual matching.
• LCMK combines the advantages of both BoV and kernel-based metrics, enhancing effectiveness.
• Extensive experiments on benchmark datasets (15-Scenes, Caltech101/256, PASCAL VOC 2007/2011) validate the method's effectiveness.

Conclusions:

• The LCMK method provides an effective and computationally efficient solution for visual similarity measurement.
• This approach offers a scalable solution for large-scale image set matching.
• The LCMK method demonstrates superior performance compared to existing techniques.