Hyperplane-Based Image Similarity Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision systems face a mismatch between training and testing phases due to the use of different loss functions, leading to suboptimal direct comparison of feature vectors for image similarity, which can be inefficient and inaccurate.
Innovation Solution
The system compares feature vectors using hyperplanes defined by benchmark images, calculating similarity based on the distance to these hyperplanes, which aligns with the training loss function, and approximates the normal of the hyperplane using the mean of positive and negative feature vectors, allowing for efficient and accurate similarity determination without requiring SVM optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If direct comparison of feature vectors using L2 distance is used during testing, then computational efficiency is maintained, but accuracy deteriorates due to mismatch with training loss functions
Solution Approach 1:
The patent introduces hyperplanes as intermediary objects that mediate between the training phase (using margin-based loss functions) and the testing phase (using efficient distance measurements). Instead of directly comparing feature vectors, the system compares hyperplanes that were optimized during training, using these hyperplanes as intermediaries to achieve both accuracy and efficiency.
Solution Approach 2:
The patent transforms the comparison task from directly measuring feature vector distances to measuring distances to hyperplanes. This parameter transformation allows the system to use the same margin-based metrics during testing that were optimized during training, resolving the online/offline mismatch while maintaining computational efficiency through simple distance calculations.
2Measurement precision
If hyperplane comparison methods are used to align with training loss functions, then accuracy improves, but computational complexity increases
Solution Approach 1:
The patent extracts the essential decision boundary information from the training process in the form of hyperplane parameters (normal vectors and offsets). These extracted hyperplane parameters are then used during testing without requiring the computationally intensive SVM optimization process, achieving accurate comparisons through simple distance calculations to the pre-computed hyperplanes.
3Measurement precision
If SVM optimization is used to determine hyperplanes, then separation accuracy improves, but computational efficiency deteriorates
Solution Approach 1:
The patent performs the computationally intensive hyperplane determination during the offline training phase, when computational resources are abundant. The resulting hyperplane parameters are stored and reused during online testing, eliminating the need for repeated SVM optimization. This preliminary action separates the heavy computation from the real-time operation, achieving both accuracy and efficiency.
Data Source
AI summary
An image processing system includes a memory to store data indicative of benchmark feature vectors of benchmark images, an input interface to accept data indicative of a first feature vector of a first image and a second feature vector of a second image, and an output interface to render a similarity value between the first and the second images. The system includes a processor to determine the similarity using a first hyperplane separating the benchmark feature vectors from the first feature vector and a second hyperplane separating the benchmark feature vectors from the second feature vector. The processor determines a first normal vector to the first hyperplane as the difference between the first feature vector and the mean of the benchmark feature vectors. The processor determines an offset for the first hyperplane as the average of the maximum inner product of the benchmark feature vectors with the first normal vector and the inner product of the first feature vector with the first normal vector. The processor determines the similarity value as a function of a sum of a signed distance of the second feature vector to the first hyperplane and a signed distance of the first feature vector to the second hyperplane.