This project is a proof-of-concept project for the on-screen translation for my game launcher: https://github.com/pat266/game-launcher
- Using EmguCV for text detection (where the text is at by drawing bounding rectangle around it) and Onnx model for text extraction gives the best result
- The Onnx model was pre-trained for recognizing Chinese text.
- It gives much faster result.
- However, it uses 2x or 3x more memory than IronOCR as the image size increases
- The translator has a delay
- 5-8 seconds in the images below
- varied depending on the number of text instances to translate
- The EmguCV + Onnx process has a batch translation system
- reduced time from 9 seconds to 2 seconds for the example below
- Basically, instead of translating one-by-one, we combine them all together (limit by certain number of characters due to GT limitation), then separate them after translation
Notice:
Copy the models folder to ```\Image_Translator\bin\Debug (or Release)``` (there will be prompt to the exact location if you have not done so)
Debug\models\
├── angle_net.onnx
├── crnn_lite_lstm.onnx
├── dbnet.onnx
└── keys.txt
EmguCVfor Text Detection- Draw Bounding Rectangle around the text
- Change values (i.e.
ar,brect.Width,brect.Height, etc.) located inGetBoudingRectangles()inForm1.csto suit your image
IronOCRfor Text Extraction- Not using Tesseract since IronOCR gives higher accuracy
Onnxfor Text Extraction- Performs much better than
IronOCRat a faster time
- Performs much better than
Google Translatefor Text Translation
- Emgu.CV (4.4.0.4099)
- Emgu.CV.Bitmap (4.4.0.4099)
- Emgu.CV.runtime.windows (4.4.0.4099)
- IronOcr (2022.3.0)
- IronOcr.Languages.Chinese (2020.11.2)
- System.Drawing.Common (4.7.0)
- GTranslate (2.1.0)
- Installing this from Nuget also installs a lot more lib
- Microsoft.ML.OnnxRuntime (1.12.0)
- Microsoft.ML.OnnxRuntime.Managed (1.12.0)
