CNN Classifier Training with Dual Labels for Image Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification techniques using convolutional neural networks (CNNs) primarily focus on binary classification rather than ranking images based on their similarity, which limits their effectiveness in image searching applications.
Innovation Solution
An image searching apparatus that employs a CNN classifier trained using a combination of first and second correct labels. The first correct label is based on predefined classes, while the second correct label is determined by the degree of similarity among images, allowing for fine-tuning of the classifier to enhance similarity-based image retrieval.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If CNN-based classifier is trained using only first correct labels (predefined classes), then the classifier can perform basic image classification, but it cannot effectively rank images by similarity for search applications
Solution Approach 1:
The patent combines two types of correct labels (first correct labels based on predefined classes and second correct labels based on similarity) to train the CNN classifier. This merging of multiple labeling approaches enables the classifier to perform both basic classification and similarity-based ranking, resolving the contradiction between adaptability and measurement precision.
Solution Approach 2:
The patent introduces a new dimension of labeling by adding second correct labels that represent similarity degrees, beyond the traditional first correct labels based on predefined classes. This additional dimensional information enables the classifier to rank images by similarity while maintaining its classification capability.
2Productivity
If the classifier is trained to focus on similarity ranking, then image search effectiveness improves, but the training complexity and computational resources increase
Solution Approach 1:
The patent performs preliminary action by generating second correct labels based on similarity calculations before the main training process. This preprocessing step organizes the similarity information in advance, which simplifies the subsequent training process and reduces computational complexity during training while maintaining high search efficiency.
3Measurement precision
If multiple correct labels are attached to each image for comprehensive training, then the classifier achieves better similarity-based search performance, but the data processing and storage requirements increase
Solution Approach 1:
The patent applies local quality by attaching different types of correct labels (first and second correct labels) to different aspects of the same image data. The first correct labels provide class information while the second correct labels provide similarity information. This localized differentiation allows the system to process and utilize each type of label appropriately, achieving high accuracy without overwhelming data processing requirements.
Data Source
AI summary
A training apparatus includes a processor which acquires an image to which identification information of a subject included in the image is attached as a first correct label the image being included in an image dataset for training and being an image for use in supervised training. The processor further attaches, to the acquired image, classification information based on a feature amount of the acquired image as a second correct label, trains a classifier using the acquired image and the second correct label attached thereto, and updates training content used in the training using the acquired image and the first correct label after training the classifier.