Building Kakao's Automated Harmful Image Classification System: Model Optimization and Cognitive Psychology-Based Labeling Tool Development

Kakao built a harmful image classification system to overcome the limitations of manual review for rapidly growing image content and ensure judgment consistency. Initially, they attempted similar image search and CNN-based classification, but refined the architecture and operational tools to satisfy both efficiency and accuracy in the actual service environment.

Evolution of Model Architecture

In the initial exploration phase, Approximate KNN was used to automatically filter similar images that had already been judged, focusing on reducing the scope of manual review. Subsequently, CNN models such as Inception-ResNet-v2 were experimented with to improve classification performance, but limitations in capturing Global Context were discovered.

To address this, the Swin Transformer was introduced. Through Window-based Self-Attention and Shifted Window mechanisms, it was possible to understand relationships across the entire image while maintaining computational efficiency. As a result, Robustness was secured, ensuring stable prediction performance even during production server deployment, and it particularly contributed to reducing False Positive cases.

Innovation in Data Quality and Labeling Tools

Based on the judgment that the upper limit of model performance depends on data quality, the labeling process was significantly improved. First, an Active Learning strategy was applied to selectively label data that the model predicted with uncertainty. Considering computing resources, Uncertainty Sampling was adopted instead of QBC (Query-by-Committee), which ensembles multiple models, to enhance operational efficiency.

Additionally, principles of cognitive psychology were applied to redesign the UI/UX of the labeling tool. Considering the limits of human working memory, unnecessary information elements on the screen were removed, and the interface was changed to be image-centric. By using Grad-CAM to visually highlight areas the AI focused on, and providing similar images and text interpretation information to assist labelers' judgments, cognitive load was reduced and judgment consistency was ensured.