Korea's Top Language Learning App Uses Warboy to Implement OCR AI

ePopSoft, Korea's largest English learning app developer, unveiled a new 'Smart Dictionary' feature using FuriosaAI's Gen 1 Vision NPU. This feature analyzes text in photos taken with a smartphone to provide accurate translations that fit the context and idiomatic expressions.

The service runs on a pipeline combining CRAFT (text detection) and ResNet-CTC (text recognition) models based on Furiosa's NPU in a cloud environment. The NPU delivers peak performance of 64TOPS (INT8) with a low design power (TDP) of 40W~60W.

With the introduction of the new OCR system, SayVoca improved accuracy by 84% compared to the existing solution. In particular, the key is accurately translating words like 'take,' whose meaning changes depending on context, by grasping the flow of the entire paragraph.

Additionally, the following technical challenges that can occur in real-world environments were resolved:

• Recognition of complex layouts such as multi-column formats, captions, and diagrams
• Handling tilted or distorted text
• Dealing with noise (artifacts) such as page stains
• Maintaining low latency on high-resolution images