Kernel Density Estimation for Feature Vector Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing distance estimation methods for feature vector sets, particularly in computer vision applications, often fail to accurately model distributions and provide strong discriminative ability due to insufficient training data and poor fitting of probabilistic models.
Innovation Solution
The proposed solution employs a non-parametric model with a kernel function optimized using labeled training data, specifically utilizing kernel density estimation and modified neighborhood component analysis to calculate distances like Kullback-Leibler divergence between feature vector sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If parametric models are used for distance estimation, then the model structure is simple, but the model fitting accuracy deteriorates due to insufficient training data
Solution Approach 1:
The patent transitions from parametric models with fixed distribution assumptions to non-parametric models that adapt distribution parameters from data. By using kernel density estimation with optimized bandwidth parameters and transformation matrices learned from training data, the system achieves better distribution fitting without requiring simple parametric assumptions.
2Productivity
If conventional distance estimation methods are used, then the computational process is simple, but the discriminative ability deteriorates
Solution Approach 1:
The patent performs preliminary optimization of kernel functions and transformation matrices using labeled training data before actual distance estimation. This pre-computation of optimal parameters through modified neighborhood component analysis enables both high discriminative ability and efficient computation during deployment, as the heavy optimization work is done in advance.
Solution Approach 2:
The system uses labeled training data to provide feedback for optimizing kernel function parameters and transformation matrices. This feedback loop through modified neighborhood component analysis adjusts the model parameters to maximize classification performance, thereby enhancing discriminative ability while maintaining computational efficiency.
3Measurement precision
If non-parametric models with optimized kernel functions are used, then the distribution fitting improves, but the computational complexity increases
Solution Approach 1:
The patent performs the computationally intensive kernel function optimization and transformation matrix learning during the training phase using labeled data. By completing these complex computations in advance, the system achieves accurate distribution fitting while keeping the actual distance estimation process computationally efficient, as the optimization work is done beforehand.
Data Source
AI summary
Apparatus, method, computer program product and computer readable medium are disclosed for distance estimation between samples. The method includes: modeling the distribution of each of two feature vector sets by a non-parametric model; and calculating the distance of the two distributions, wherein a kernel function is used in the non-parametric model, the kernel function is optimized based on labeled training data, the first feature vector set includes a plurality of feature vectors extracted from a sample, and the second feature vector set includes a plurality of feature vectors extracted from another sample.