Image Classification Apparatus Merging Rare Classes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification technologies face challenges in accurately classifying images when there are classes with insufficient training images, leading to erroneous classification results.
Innovation Solution
An information processing apparatus is designed to acquire the number of training images for each class, identify classes with fewer images than a threshold, combine these classes with adjacent classes, and use a determiner trained on the combined classes to classify input images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training images are collected for each class to improve classification accuracy, then determination accuracy is improved, but classes with abnormal states have insufficient training images leading to erroneous classification
Solution Approach 1:
The patent combines classes with insufficient training images (abnormal states) with adjacent normal classes to create composite training datasets. This merging allows the determiner to learn from sufficient training data while still covering abnormal states, thereby resolving the contradiction between needing sufficient training images and maintaining classification accuracy for rare abnormal classes.
2Reliability
If a determiner is trained with sufficient training images for all classes, then classification accuracy is improved, but obtaining sufficient training images for abnormal classes becomes more difficult
Solution Approach 1:
The patent merges abnormal classes with adjacent normal classes to create composite training categories. This approach makes it easier to obtain sufficient training images because normal class images are abundant, while still enabling the determiner to reliably identify abnormal states through the combined training framework.
Solution Approach 2:
The patent introduces an intermediate processing step where classes are combined before training. This intermediary approach allows the system to bypass the difficulty of directly collecting sufficient abnormal class images by using normal class images as intermediaries in the training process.
3Reliability
If classes are combined to compensate for insufficient training images, then classification reliability is improved, but the complexity of class management increases
Solution Approach 1:
The patent segments the class combination process into distinct steps: identifying classes with insufficient images, selecting adjacent classes for combination, creating composite training datasets, and training the determiner. This segmentation makes the complex process more manageable and systematic.
Solution Approach 2:
The patent changes the parameter of class granularity by combining fine-grained abnormal classes into coarser composite classes for training purposes. This parameter change simplifies class management while maintaining the ability to detect abnormal states through the combined categories.
Data Source
AI summary
An information processing apparatus includes an acquisition unit configured to acquire the number of training images including an object belonging to one of a plurality of classes having an order relation, the acquisition unit acquiring the number of training images for each of the plurality of classes, a first determination unit configured to determine, as a combination target class to be combined, a class with the number of training images that is smaller than a threshold value among the plurality of classes, a combination unit configured to combine the combination target class with an adjacent class in the order relation, and a second determination unit configured to determine a class of an object included in an input image using a determiner trained using the training image based on a combination class combined by the combination unit.