Text recognition for seven segment display using keras-ocr.

OCR for seven segment display using tensorflow lite. The trained keras model is converted to tensorflow lite model model_float16.tflite

Setup

1. Clone the repo locally

   git clone https://github.com/renjithsasidharan/seven-segment-ocr
   

2. Install dependencies

   pip install requirements.txt
   

Training Data

Training data is present in data/ocr_training_data

data/ocr_training_data
├── 00498afd-2925-45ef-bfb3-8c55204ace42.png
├── 011fd0b2-837c-424c-a71b-a2a92305a532.png
├── 04976571-ac6b-425f-9474-3b95a4fb9613.png
├── 068e4ff1-acf0-43e7-b052-7a59a9c3f87b.png

Ground truth file is data/ocr_training_data/gt.txt

Create more training data

You can create more training data by first creating annotations and then creating training data from annotations

Create annotations

1. Add images to data/annotations folder.
2. Install annotation tool PPOCRLabel from https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.4/PPOCRLabel
3. Open PPOCRLabel and start annotating new images.
   1. Create RectBox around the reading
   2. Add the reading in Recognition Result
   3. Save annotations

Create Training data from annotations

From seven-segment-ocr directory run the command

python create_ocr_training_data_from_annotations.py

This will append the training data to data/ocr_training_data

How to train the model using google Colab

Train the model using jupyter notebook keras_ocr_7_seg.ipynb

1. Upload the notebook keras_ocr_7_seg.ipynb to google Colab
2. Create archive of folder data/ocr_training_data.
3. Upload ocr_training_data.zip to google Colab
4. Run the training
5. Once the training is over, your trained model meeter_rec_float16.tflite will be automatically downloaded to your machine from Colab

Evaluation

To run prediction on an image using tensorflow lite

python predict.py --image test/1.png --model model_float16.tflite

To run predictions on all images in directory

python eval_tflite.py --eval_dir /data/ocr_training_data

Here are some results of ocr extraction:

Confusion matrix