Given an input Image we need to predict the Text in the Image with a reasonable accuracy >80% (Exact match with the actual Text Labels) and should have a good letter match accuracy.
- Accuracy
- Letter Accuracy
- CTC Loss
The Model used is Convolutional Recurrent Neural Network(CRNN) which is end-to-end trainable and its architecture consists of three parts:
- convolutional layers, which extract a feature sequence from the input image;
- recurrentlayers, which predict a label distribution for each frame;
- transcription layer, which translates the per-frame predictions into the final label sequence
Please check my iPython Notebook for detailed analysis and implementation of this project.
I have written a Medium Blog giving a detailed explaination of the implementation. Do feel free to read it.
https://medium.com/analytics-vidhya/image-text-recognition-738a368368f5