Abstract
Educational platforms in multilingual regions face challenges when educators need to locate equivalent or similar diagram content across different languages. Traditional text-based search is limited when educational diagrams exist in multiple languages, such as Russian, Kazakh, and English, with identical visual structures but different textual annotations. This paper presents two multimodal diagram search methods implemented within Kazakhstan’s professional facilitator network. The baseline approach utilises OpenCV feature extraction, combining colour histograms, texture descriptors, edge detection, and shape characteristics to facilitate rapid visual similarity detection. The enhanced method integrates ResNet-50 deep features with multilingual optical character recognition processing. Both methods operate within an active educational platform receiving weekly diagram submissions from practicing educators. Evaluation using authentic educational content shows the baseline method achieves processing times of 1.2–1.8 seconds, suitable for real-time collaboration, while the enhanced method achieves cross-lingual similarity retrieval with higher computational requirements. The baseline method performs well on workflow diagrams and process charts, while the enhanced method shows improved accuracy for conceptual illustrations requiring semantic understanding. The deployed system demonstrates that combining traditional computer vision with deep learning features addresses the practical requirements of multilingual educational collaboration through visual similarity detection rather than text matching.