Equation Detector(For CV Term Project)
TO detect a Equation from an image and export the result as a Latex form
Difference from other Detector
It detects the form directly from the image, not the inkML. (InkML contains the order of the drawing.)
-
Make a single Symbol Detector using several kinds of Networks. (Caution : We should unify the image size in order to inference the class correctly)
-
After the performance of the single symbol classification is verified, put the equation image to the model and compare the feature.
-
Use LSTM to train the order of the symbols.
Default input image size 96 * 480
INKML images are reshaped to 80 * 448, and will be put in the center of the frame.
Test Set for Equation : Testdata in 2014 CHORME(988 items)
- Edit loss plot(need to reduce bouncing graph)
- Encode the latex code according to the dictionary
- Check the encoding algorithm by making a decoder
- Keep updating dictionary, check for errors
- Implement loading model from pt file
- Save optimizer parameters
- Add resume functions
- Create Dataloader for RNN input/output
- (OPTIONAL) Hierarchical softmax
- (OPTIONAL) More data for symbol classification
- Post Processing in LSTM
- Load various cnn models in CRNN