Аналіз питання перекладу тексту на зображеннях на основі мультимодальних великих мовних моделей

Abstract

Traditional image translation typically relies on sequential OCR and machine translation, which often fail to account for visual layouts and context. This research evaluates multimodal large language models in simultaneously processing textual and visual data to ensure accurate, context-aware translations while preserving document topology. The study demonstrates that these models automate 80–90% of visual content localisation by maintaining structural integrity across various formats. Despite these advancements, human verification remains necessary for low-quality images to prevent errors. Future work in this field will focus on technologies like visual font transfer to achieve seamless, design-consistent localisation