Image Identification Model Training With Similarity-Aware Pseudo Classes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image identification models struggle with generating effective training data for product identification, leading to overlearning and misidentification due to a lack of consideration for the similarity between original and converted images.
Innovation Solution
An information processing apparatus and method that generates converted images from original images using image conversion parameters, assigns pseudo classes, and adjusts an objective function to reflect similarity between original and converted images, enabling the image identification model to learn effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image conversion process is applied to generate training data, then the quantity of training data is improved, but the identification accuracy deteriorates due to overlearning
Solution Approach 1:
The patent applies parameter changes by modifying the objective function to include correction terms that account for image similarity. Specifically, the loss function is adjusted to penalize excessive separation between original and converted images, allowing the model to learn from converted images without overlearning. This resolves the contradiction by transforming the training approach while maintaining data quantity benefits.
Solution Approach 2:
The patent implements feedback by incorporating similarity-based correction terms into the loss function. The model receives feedback about the similarity between original and converted images, adjusting its learning to prevent over-separation. This feedback mechanism allows the system to maintain high identification accuracy while utilizing converted images for training.
2Adaptability or versatility
If converted images are assigned pseudo classes different from original classes, then the model learns to distinguish different forms, but the similarity information is lost leading to misidentification
Solution Approach 1:
The patent changes the parameter formulation of the loss function to include similarity correction terms. Instead of treating converted images as completely different classes, the objective function incorporates terms that preserve similarity information, allowing the model to learn both distinction and similarity simultaneously.
Solution Approach 2:
The patent creates a composite learning objective that combines multiple components: standard classification loss and similarity-based correction terms. This composite approach allows the model to learn both the ability to distinguish different product forms and maintain awareness of their similarities, preventing misidentification.
Data Source
AI summary
Provided is an information processing apparatus causing an image identification model to perform learning in consideration between a degree of similarity between an original image and a converted image. The information processing apparatus includes at least one processor configured to execute: a training data generating process of generating converted images from original images using image conversion parameters and giving, to the generated converted images, pseudo classes corresponding to classes of the original images and the image conversion parameters, respectively; a loss-for-training determining process of generating an objective function in which correction terms each given in consideration of a degree of similarity between a corresponding one of the classes and a corresponding one of the pseudo classes is reflected, the correction terms being determined for the respective image conversion parameters; and a learning process of causing an image identification model to perform learning using the converted images and the objective function.