krockema / edocr

A packaged OCR system for mechanical engineering drawings based on keras-ocr

eDOCr

eDOCr is a packaged version of keras-ocr that facilitates end-to-end digitization of mechanical EDs. Developed for Windows OS and using Python as the primary programming language. The implementation is discussed in the researh paper Optical character recognition on engineering drawings to achieve automation in production quality control

Getting Started

Installation

eDOCr supports Python >= 3.6 and TensorFlow >= 2.0.0. The test environment I used is Python = 3.9 and Tensorflow = 2.10.1. Install your prefered distribution platform, Anaconda is recommended. Open Anaconda Prompt and type the following commands:

conda create -n edocr python=3.9 -y 
conda activate edocr

# To install from PyPi
conda install pip
pip install eDOCr

# To install from Source
cd path/to/your/folder
git clone https://github.com/javvi51/eDOCr
cd eDOCr
pip install -r requirements.txt
pip install .

Using

There are two ways of using eDOCr: from terminal and from your own python file.

From Terminal:

We need to locate the ocr_it.py file in our system. If you have installed using pip, it will probably come in C:\Users\YOUR_USER\.conda\envs\edocr\Lib\site-packages\eDOCr\ocr_it.py. If you have installed from source, it will be at your selected folder. All you need to do is:

python PATH/TO/YOUR/FOLDER/eDOCr/ocr_it.py PATH/TO/YOUR/DRAWING/my_drawing.pdf

Additional commands you can use are:

# Specify the destination path. By default, it is the path you are running your code from.
--dest-folder PATH/TO/YOUR/DESTINATION/FOLDER
# Does the drawing have watermark you want to remove? By default, it is not considered.
--water
# Advance Setting: Set a custom threshold distance (in px.) for grouping detections. Default is 20px.
--cluster 25

From your own python file

More customization is possible using your own python file, such as selecting a different model, alphabet or changing colors.

# Importing packages
import os
from eDOCr import tools
import cv2
import string
from skimage import io

# Loading image and destination file
dest_DIR = 'tests/test_Results'
file_path = 'tests/test_samples/Candle_holder.jpg'
filename = os.path.splitext(os.path.basename(file_path))[0]
img = cv2.imread(file_path)

# Selecting alphabet and model (Note that alphabet and alphabet model need to match)
GDT_symbols = '⏤⏥○⌭⌒⌓⏊∠⫽⌯⌖◎↗⌰'
FCF_symbols = 'ⒺⒻⓁⓂⓅⓈⓉⓊ'
Extra = '(),.+-±:/°"⌀'

alphabet_dimensions = string.digits + 'AaBCDRGHhMmnx' + Extra
model_dimensions = 'eDOCr/keras_ocr_models/models/recognizer_dimensions.h5'
alphabet_infoblock = string.digits+string.ascii_letters+',.:-/'
model_infoblock = 'eDOCr/keras_ocr_models/models/recognizer_infoblock.h5'
alphabet_gdts = string.digits + ',.⌀ABCD' + GDT_symbols
model_gdts = 'eDOCr/keras_ocr_models/models/recognizer_gdts.h5'

# Selecting personalized color palette and cluster setting
color_palette = {'infoblock': (180, 220, 250), 'gdts': (94, 204, 243), 'dimensions': (93, 206, 175), 'frame': (167, 234, 82), 'flag': (241, 65, 36)}
cluster_t = 20

# eDOCr functions
class_list, img_boxes = tools.box_tree.findrect(img)
boxes_infoblock, gdt_boxes, cl_frame, process_img = tools.img_process.process_rect(class_list, img)
io.imsave(os.path.join(dest_DIR, filename + '_process.jpg'), process_img)

infoblock_dict = tools.pipeline_infoblock.read_infoblocks(boxes_infoblock, img, alphabet_infoblock, model_infoblock)
gdt_dict = tools.pipeline_gdts.read_gdtbox1(gdt_boxes, alphabet_gdts, model_gdts, alphabet_dimensions, model_dimensions)
 
process_img = os.path.join(dest_DIR, filename + '_process.jpg')

dimension_dict = tools.pipeline_dimensions.read_dimensions(process_img, alphabet_dimensions, model_dimensions, cluster_t)
mask_img = tools.output.mask_the_drawing(img, infoblock_dict, gdt_dict, dimension_dict, cl_frame, color_palette)

# Record the results
io.imsave(os.path.join(dest_DIR, filename + '_boxes.jpg'), img_boxes)
io.imsave(os.path.join(dest_DIR, filename + '_mask.jpg'), mask_img)
tools.output.record_data(dest_DIR, filename, infoblock_dict, gdt_dict, dimension_dict)

Training a model on a custom alphabet

Fonts are not loaded if installing from pip. To train new models, please install from source.

To train a model in a custom alphabet, a python file is provided, so that the only steps needed are:

# Importing Packages
import os
import string
from eDOCr import keras_ocr
from eDOCr.keras_ocr_models import train_recognizer

# Fixing paths and alphabet
DIR = os.getcwd()
recognizer_basepath = os.path.join(DIR, 'eDOCr/Keras_OCR_models/models')
data_dir = './tests'
alphabet = string.digits + 'AaBCDRGHhMmnx' + '().,+-±:/°"⌀'

# Number of autogenerated samples
samples = 10000

# Load white backgrounds and fonts
backgrounds = []
for i in range(0, samples):
    backgrounds.append(os.path.join('./eDOCr/Keras_OCR_models/backgrounds/0.jpg'))

fonts = []
for i in os.listdir(os.path.join(DIR, 'eDOCr/Keras_OCR_models/fonts')):
    fonts.append(os.path.join('./eDOCr/Keras_OCR_models/fonts', i))

# Choose a pretrained model if you like

pretrained_model = None 
#pretrained_model = os.path.join(recognizer_basepath,'recognizer_dimensions.h5')

# Start Training 
train_recognizer.generate_n_train(alphabet, backgrounds, fonts, recognizer_basepath=recognizer_basepath, pretrained_model=pretrained_model)

This work was supported by Vinnova (Sweden) under grant 2021-02481, iPROD project. The authors would like to thank Vinnova for making this research project possible.