Hand Written Recognition Project

Task 1 and 2

First, ensure to have Python version 3.8. Then, install the required packages using the following command:

pip install -r requirements.txt

Test Image

Keep your images in one folder. Our model is only designed for binary image.

You can pass the folder path of images by using --image_path "Path/to/Folder"

Character segmentation using pretrained model

We have 4 different models. The pretrained models can be found here.. We recommend to download the InceptionResNetV2, since it yielded the highest accuracy on the test data. Then, pass the path of the classifier using --classifier_path "Path/of/model"

Use the corresponding flag from the following list based on the model you are using for results:

1. scratch implemented CNN (flag = 0)
2. InceptionV3 (flag = 1)
3. ResNet50 (flag = 2)
4. InceptionResNetV2 (flag = 3)
   

Example:

python main.py --image_path "Path/to/Folder" --classifier_path "Path/to/model" --flag 3

Training Classifier

This part is optional and is not required for evaluating the pipeline, as the classifier is already trained. If you wish to retrain the classifier you can run:

The script includes 4 different models, which can be chosen by adjusting the flag

1. scratch implemented CNN (flag = 0)
2. InceptionV3 (flag = 1)
3. ResNet50 (flag = 2)
4. InceptionResNetV2 (flag = 3)
   

Furthermore, the number of augmented pictures n can be modified by selecting number_of_augmentations_per_class = n. The current parameters are the parameters which were used to train our classifier (InceptionResNetV2, no augmentation). After running the script, the model is automatically saved in Model_dir. It is also possible to pass model directory as argument.

Example:

python classifier_pipeline.py --flag 0 --model_dir "models" --epochs 50 --aug_no 0

Sliding Window

This approach is also optional and not our best performing approach, however, we include it due to matters of completeness. For evaluating the accuracy of our pipeline please follow the stepts in the section Character segmentation using pretrained model above.

Again, after downloading the pretrained classifier you can choose the classifier by the aforementioned flags. Example:

python sliding_window.py --image_path "Path/to/images" --classifier_path "Path/to/model" --flag 3