Image Recognition Score Normalization via Logistic Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition systems struggle with comparing analyses across numerous concepts due to non-comparable raw scores, as scores for different concepts are not directly equivalent, leading to inconsistent results.
Innovation Solution
A system that calculates raw scores for media items using a classifier model, ranks them based on a target concept, and applies a normalization process through a logistic regression model to convert raw scores into normalized scores, allowing for comparison across concepts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If raw scores are used directly for comparing concept likelihoods across media items, then the classification process is simple and fast, but the results are inconsistent and non-comparable across different concepts
Solution Approach 1:
The patent transforms the raw score parameter into a normalized probability parameter through logistic regression. The raw score (unbounded real number) is changed into a normalized score (probability between 0 and 1), enabling meaningful comparisons across different concepts while maintaining the underlying classification information.
Solution Approach 2:
The patent introduces logistic regression as an intermediary transformation layer between the classifier model and the final comparison. This intermediary process converts raw scores into normalized probabilities, serving as a bridge that enables cross-concept comparability without directly modifying the original classification outputs.
2Reliability
If a normalization process is applied to make scores comparable across concepts, then consistent results are achieved, but the processing time and computational complexity increase
Solution Approach 1:
The patent performs preliminary normalization by pre-processing the training data and fitting the logistic regression model during the training phase. Once trained, the normalization parameters (coefficients and threshold) are stored and applied efficiently during inference, reducing the time penalty during actual media item processing.
Solution Approach 2:
The patent replaces complex post-processing comparison mechanisms with a simplified probability-based system. By substituting the need for concept-specific score interpretation with universal probability values, the system achieves consistency without requiring complex adjustment mechanisms during runtime.
3Measurement precision
If logistic regression is used for normalization, then accurate probability values are obtained for comparison, but the model training complexity increases
Solution Approach 1:
The patent segments the overall training process into two independent stages: (1) training the base classifier model to generate raw scores, and (2) training the logistic regression model to normalize those scores. This segmentation allows each model to be trained separately on its specific task, simplifying the overall training complexity while maintaining accuracy.
Solution Approach 2:
The patent uses feedback from the base classifier's raw scores to train the logistic regression model. The raw scores serve as input features for the normalization model, creating a feedback loop where the output of one model becomes the input of the next, enabling accurate probability calibration based on actual classification performance.
Data Source
AI summary
Systems, methods, and non-transitory computer-readable media can calculate raw scores for a plurality of media items based on a classifier model and a target concept. The plurality of media items are ranked based on the raw scores. A review set of the plurality of media items is determined, the review set comprising a subset of the plurality of media items. Each of the media items of the review set is associated with a content depiction determination. A normalized score formula is calculated based on the raw scores and the content depiction determinations for the media items of the review set.