Feature Extractor Loss Function for Image Authentication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image authentication techniques face challenges in improving accuracy, particularly when pair similarities are close to or far from a threshold, leading to suboptimal learning effects.
Innovation Solution
A learning apparatus that calculates feature vectors from images using a feature extractor, focuses on pair similarities near a threshold, and employs a loss function with a gradient that is large near the threshold but decreases as similarity deviates from it, to enhance authentication accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all pairs are used for learning regardless of similarity, then learning coverage is comprehensive, but learning efficiency is low and accuracy improvement is limited
Solution Approach 1:
The patent applies local quality by differentiating the treatment of pairs based on their similarity characteristics. Pairs near the threshold receive focused learning attention with higher gradient magnitudes, while pairs far from the threshold receive reduced learning emphasis. This selective approach concentrates computational resources on the most critical cases (pairs near threshold) that最需要improvement, thereby resolving the contradiction between comprehensive learning coverage and learning efficiency.
2Reliability
If learning focuses only on pairs near threshold, then authentication accuracy improves, but pairs with much higher or lower similarity are neglected
Solution Approach 1:
The patent implements dynamics by making the learning emphasis adaptive rather than static. The gradient magnitude dynamically adjusts based on the pair's similarity value relative to the threshold: pairs near the threshold receive larger gradients for focused learning, while pairs far from the threshold receive smaller but non-zero gradients. This dynamic approach ensures that the model adapts its learning focus to the specific characteristics of each pair, resolving the contradiction between focused accuracy improvement and comprehensive learning coverage.
3Reliability
If a standard loss function is used, then learning is stable, but learning effectiveness for threshold-critical pairs is insufficient
Solution Approach 1:
The patent applies parameter changes by modifying the loss function's gradient characteristics based on the pair's similarity parameter. Specifically, the gradient magnitude is adjusted as a function of the similarity value: pairs with similarity close to the threshold generate larger gradient magnitudes, while pairs with similarity far from the threshold generate smaller gradient magnitudes. This parameter-based adjustment makes the loss function more precise in guiding learning for threshold-critical pairs, resolving the contradiction between learning stability and learning effectiveness.
Data Source
AI summary
An apparatus calculates a positive pair similarity between a first feature vector of an image and a second feature vector associated with the same label as a label of the image, calculates a negative pair similarity between the first feature vector and a third feature vector associated with a label different from the label of the image, calculates a loss value with respect to the positive pair similarity lower than a threshold, calculates a loss value with respect to the negative pair similarity higher than the threshold, and learns a parameter of an feature extractor that decreases the loss values. For calculation one of the loss values, a loss function including a function such that an absolute value of a gradient is larger near a predetermined threshold and is non-zero but smaller even at a point away from the threshold is used.