Exploiting Web Images for Multi-Output Classification: From Category to Subcategories.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Image classification accuracy improves with hierarchical categorization (categories and subcategories).
• Current subcategorization methods require extensive expert input and labeled datasets, which are costly and time-consuming.
• Automated approaches are needed to overcome the limitations of traditional subcategorization techniques.

Purpose of the Study:

• To develop an automated method for image subcategorization that eliminates the need for expert knowledge and labeled images.
• To enhance image classification performance through effective category and subcategory assignment.
• To propose a unified framework for simultaneously removing outlier images and learning optimal classification models.

Main Methods:

• Candidate subcategory labels are extracted from untagged text corpora.
• Subcategory labels are refined by calculating their relevance to the target category.
• A unified problem formulation integrates outlier image removal and optimal classification model learning, combining subcategory classifiers to form category classifiers.

Main Results:

• The proposed method successfully identifies and purifies relevant subcategory labels without expert intervention.
• The unified approach effectively suppresses noise from outlier images and improves classification model performance.
• Experimental results demonstrate superior performance in both image categorization and subcategorization tasks compared to existing methods.

Conclusions:

• The developed approach offers a significant advancement in automated image subcategorization.
• It reduces the dependency on manual labeling and expert knowledge, making subcategorization more efficient and scalable.
• The method holds promise for improving the accuracy and efficiency of large-scale image classification systems.