Image Classifier Group Training for Size and Brightness Errors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to improve classification accuracy between categories of objects that are easily misidentified under specific conditions, such as small-sized objects or low brightness images, leading to misidentifications like a small airplane being mistaken for a bird or a dog in low brightness as a cat.
Innovation Solution
An information processing apparatus that generates groups of categories likely to be misidentified under specific attributes by aggregating attribute parameters and identification results, and trains the model using a group softmax function to enhance classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional training methods are used for category classification, then general classification performance is maintained, but classification accuracy between categories easily misidentified under specific conditions deteriorates
Solution Approach 1:
The patent segments categories into groups based on their susceptibility to misidentification under specific conditions. Categories are divided into a first group (easily misidentified) and a second group (not easily misidentified), allowing differential training strategies to be applied to each group, thereby improving classification accuracy for problematic categories without compromising overall performance
Solution Approach 2:
The patent applies local quality by implementing condition-specific training for categories in the first group. The training process is customized based on attributes such as size and brightness, where categories are trained under specific condition settings that match their misidentification patterns. This localized training approach improves accuracy for specific problematic cases while maintaining efficient general training for other categories
2Reliability
If training data is divided into groups by occurrence frequency, then classifier imbalance is suppressed, but classification accuracy between categories under specific attribute conditions remains insufficient
Solution Approach 1:
The patent introduces dynamic training by adjusting training conditions based on attribute values. The training process is not static but adapts to specific conditions such as size ranges and brightness levels. Categories are trained under dynamically selected condition settings that match their misidentification patterns, allowing the system to handle both class imbalance and condition-specific accuracy requirements
Data Source
AI summary
An information processing apparatus for training a model that identifies a category of an object included in an image includes at least one processor and at least one memory that is in communication with the at least one processor. The at least one memory stores instructions for causing the at least one processor and the at least one memory to acquire an attribute from the image, acquire information about a group of categories easily misidentified with each other under a specific attribute condition, generate a group including a plurality of categories when the model is trained based on the attribute and the information about the group, and train the model based on an identification result generated by identifying the category of the object included in the image using the model, and the group of the categories.