Refining Image Annotations via Confidence Threshold Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image search systems face challenges in accurately identifying high-confidence labels and images for training image relevance models, leading to decreased precision due to inclusion of labels and images that are difficult to correlate, such as overly specific or general annotations.
Innovation Solution
The method involves determining confidence values for image labels based on precision and frequency thresholds, identifying high-confidence labels and images, and excluding those that do not meet these thresholds from the training set to refine the image relevance model, thereby improving the accuracy of image annotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If all available image labels are included in the training set, then the quantity of training data increases, but the precision of image annotations decreases due to inclusion of low-confidence labels
Solution Approach 1:
The patent segments the training image set into high-confidence images and low-confidence images based on confidence value thresholds. This segmentation allows the system to selectively use only high-confidence images for training, thereby maintaining annotation precision while still utilizing a substantial portion of the available data.
Solution Approach 2:
The patent applies local quality by assigning different confidence values to different labels within the training set. Each image label is evaluated individually, and only those meeting the confidence threshold are retained. This ensures that each component of the training data meets a quality standard, preventing dilution of overall annotation precision.
2Measurement precision
If confidence thresholds are applied to filter labels, then the precision of annotations improves, but the quantity of training images decreases
Solution Approach 1:
The patent performs preliminary action by calculating confidence values for all image labels before the training process. This pre-evaluation allows the system to identify and retain only high-confidence labels in advance, ensuring that the training set is pre-filtered for quality without requiring post-training adjustments or retraining.
Solution Approach 2:
The patent utilizes parameter changes by adjusting confidence thresholds to optimize the balance between precision and quantity. By modifying the threshold parameter, the system can control the proportion of training images retained, allowing flexible adjustment based on specific training requirements and available data volume.
3Adaptability or versatility
If image relevance model uses all labels for training, then the model's versatility increases, but the reliability of relevance determination decreases
Solution Approach 1:
The patent segments the label set into high-confidence and low-confidence categories, using only high-confidence labels for training the image relevance model. This segmentation ensures that the model learns from reliable data, improving the reliability of relevance determination while still maintaining versatility through the use of multiple high-confidence labels.
Solution Approach 2:
The patent implements feedback by using confidence values derived from user interaction data (clicks, views) to continuously refine the selection of training images. This feedback mechanism ensures that the model is trained on labels that have been validated by actual user behavior, thereby improving reliability while maintaining adaptability to user preferences.
Data Source
AI summary
Methods, systems and apparatus for refining image annotations. In one aspect, a method includes receiving, for each image in a set of images, a corresponding set of labels determined to be indicative of subject matter of the image. For each label, one or more confidence values are determined. Each confidence value is a measure of confidence that the label accurately describes the subject matter of a threshold number of respective images to which it corresponds. Labels for which each of the one or more confidence values meets a respective confidence threshold are identified as high confidence labels. For each image in the set of images, labels in its corresponding set of labels that are high confidence labels are identified. Images having a corresponding set of labels that include at least a respective threshold number of high confidence labels are identified as high confidence images.