Classifier Training via Negative Example Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification methods for detecting indecent images face high detection missing rates and error rates due to the scattered distribution of negative-example image features, leading to overlapping features with positive-example images, which results in over-fitting and inaccurate classification.
Innovation Solution
The method involves dividing negative-example sample sets into categories to determine distinct feature sets for differentiating indecent images from various types of negative-example images, using these feature sets to train multiple classifiers, thereby reducing detection missing and error rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If different types of negative-example images are collected to form a single negative-example sample set, then the sample set becomes more comprehensive, but the feature distribution becomes more scattered and overlap with positive-example features increases
Solution Approach 1:
The patent segments the single negative-example sample set into multiple category-specific negative-example sample sets (e.g., portrait images, landscape images, food images). Each category-specific set maintains homogeneous feature distributions, preventing the scattering effect that occurs when diverse negative examples are mixed. This segmentation allows comprehensive coverage of different image types while preserving clear feature boundaries for accurate classification.
2Device complexity
If a single classifier is trained with mixed negative-example images, then the training process is simpler, but the detection missing rate and error rate increase
Solution Approach 1:
The patent divides the classification task into multiple specialized classifiers, each trained on a specific category of negative examples (e.g., one classifier for portraits, another for landscapes). This segmentation enables each classifier to specialize in distinguishing positive examples from a particular type of negative example, thereby reducing detection errors that occur when a single classifier must handle all negative categories simultaneously.
Solution Approach 2:
The patent introduces a categorical dimension to the classification system by organizing negative examples into distinct categories and creating separate classifiers for each category. This dimensional organization transforms the single-classifier approach into a multi-classifier architecture, where each classifier operates in a specialized feature space optimized for its specific negative example category, thereby improving overall detection reliability.
3Loss of time
If all negative-example images are marked with the same label, then the labeling process is more efficient, but the feature distinguishability between positive and negative examples decreases
Solution Approach 1:
The patent applies local quality by assigning different labels or label characteristics to different categories of negative examples based on their specific features (e.g., portrait images receive one type of label, landscape images receive another). This localized labeling strategy enhances feature distinguishability by capturing the unique characteristics of each negative example category, enabling classifiers to better differentiate between positive examples and various types of negative examples.
Data Source
AI summary
Embodiments of the present invention provide a method and apparatus for training an image classifier. The method includes: A. dividing a set of training images for classifier training into a positive-example sample set and at least two negative-example sample sets; B. determining, for each negative-example sample set, a feature set for differentiating the positive-example sample set from the negative-example sample set; and C. performing training using each feature set determined to obtain a classifier. This invention also provides a method and apparatus for image recognition utilizing the image classifier.