Efficient HIK SVM learning for image classification.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Histograms are fundamental in image processing and computer vision.
• Histogram Intersection Kernel (HIK) and Support Vector Machine (SVM) classifiers are effective for histogram-based tasks.

Purpose of the Study:

• To develop a novel, efficient solver for HIK SVM for image classification.
• To analyze the performance and parameter sensitivity of the proposed method.

Main Methods:

• Propose Intersection Coordinate Descent (ICD), a deterministic and scalable HIK SVM solver.
• Extend ICD to train a broader family of kernels.
• Provide theoretical analysis for ICD's parameter insensitivity.

Main Results:

• ICD demonstrates significantly faster training times compared to general SVM solvers and other HIK SVM methods.
• ICD achieves comparable accuracies to existing methods.
• ICD shows robustness to the SVM C parameter, maintaining high accuracy with default settings.

Conclusions:

• ICD offers a computationally efficient and accurate solution for HIK SVM image classification.
• The parameter insensitivity of ICD is advantageous for large-scale image processing tasks where cross-validation is infeasible.