SWTloc : Stroke Width Transform Text Localizer

Header	Status
Latest Release
Downloads
Supported Python
Documentation
Open Issues
License

Description

This repo contains a python implementation structured as a python package pertaining to the text localization method as in a natural image as outlayed in the Research Paper :-

Detecting Text in Natural Scenes with Stroke Width Transform. Boris Epshtein, Eyal Ofek & Yonatan Wexler (June, 2010)

This library extends the transformation stage of the image for textual content by giving the ability to :

• Localize Letter's : through SWTImage.localizeLetters
• Localize Words's, via fusing individual Letter's : through SWTImage.localizeWords

The process flow of is depicted in the image below :

Installation

pip install swtloc

Documentation

Documentation for SWTLoc can be found at - SWTLoc Documentation

Speed Benchmarking

Below is the speed comparison between different versions of SWTLoc and their various engines. The time measured for each test image was calculated based on 10 iterations of 10 runs each. Test Images can be found in examples/images/ folder in this repository, and the code for generating the below table can be found in - Improvements-in-v2.0.0.ipynb notebook in examples/ folder.

Frequently Used Code Snippets

Performing Stroke Width Transformation

# Installation
# !pip install swtloc

# Imports
import swtloc as swt
# Image Path
imgpath = 'examples/images/test_image_5/test_img5.jpg'
# Result Path
respath = 'examples/images/test_image_5/usage_results/'
# Initializing the SWTLocalizer class with the image path
swtl = swt.SWTLocalizer(image_paths=imgpath)
# Accessing the SWTImage Object which is housing this image
swtImgObj = swtl.swtimages[0]
# Performing Stroke Width Transformation
swt_mat = swtImgObj.transformImage(text_mode='db_lf')

Localizing & Annotating Letters and Generating Crops of Letters

# Installation
# !pip install swtloc

# Imports
import swtloc as swt
from cv2 import cv2
import numpy as np
# Image Path
imgpath = 'examples/images/test_image_1/test_img1.jpg'
# Read the image
img = cv2.imread(imgpath)
# Result Path
respath = 'examples/images/test_image_1/usage_results/'
# Initializing the SWTLocalizer class with a pre loaded image
swtl = swt.SWTLocalizer(images=img)
swtImgObj = swtl.swtimages[0]
# Perform Stroke Width Transformation
swt_mat = swtImgObj.transformImage(text_mode='db_lf',
                                   maximum_angle_deviation=np.pi/2,
                                   gaussian_blurr_kernel=(11, 11),
                                   minimum_stroke_width=5,
                                   maximum_stroke_width=50,
                                   display=False)  # NOTE: Set display=True 
# Localizing Letters
localized_letters = swtImgObj.localizeLetters(minimum_pixels_per_cc=950,
                                              maximum_pixels_per_cc=5200)
letter_labels = [int(k) for k in list(localized_letters.keys())]

# Some Other Helpful Letter related functions
# # Query a single letter
from swtloc.configs import (IMAGE_ORIGINAL,
                            IMAGE_SWT_TRANSFORMED)
loc_letter, swt_loc, orig_loc = swtImgObj.getLetter(key=letter_labels[5])

# # Iterating over all the letters
# # Specifically useful for jupyter notebooks - Iterate over all
# # the letters, at the same time visualizing the localizations
letter_gen = swtImgObj.letterIterator()
loc_letter, swt_loc, orig_loc = next(letter_gen)

# # Generating a crop of a single letter on any of the available
# # image codes.
# # Crop on SWT Image
swtImgObj.saveCrop(save_path=respath,crop_of='letters',crop_key=6, crop_on=IMAGE_SWT_TRANSFORMED, crop_type='min_bbox')
# # Crop on Original Image
swtImgObj.saveCrop(save_path=respath,crop_of='letters',crop_key=6, crop_on=IMAGE_ORIGINAL, crop_type='min_bbox')

Localizing & Annotating Words and Generating Crops of Words

# Installation
# !pip install swtloc
# Imports
import swtloc as swt
# Image Path
imgpath = 'images/test_img2/test_img2.jpg'
# Result Path
respath = 'images/test_img2/usage_results/'
# Initializing the SWTLocalizer class with the image path
swtl = swt.SWTLocalizer(image_paths=imgpath)
swtImgObj = swtl.swtimages[0]
# Perform Stroke Width Transformation
swt_mat = swtImgObj.transformImage(maximum_angle_deviation=np.pi/2,
                                   gaussian_blurr_kernel=(9, 9),
                                   minimum_stroke_width=3,
                                   maximum_stroke_width=50,
                                   include_edges_in_swt=False,
                                   display=False)  # NOTE: Set display=True 

# Localizing Letters
localized_letters = swtImgObj.localizeLetters(minimum_pixels_per_cc=400,
                                              maximum_pixels_per_cc=6000,
                                              display=False)  # NOTE: Set display=True 

# Calculate and Draw Words Annotations
localized_words = swtImgObj.localizeWords(display=True)  # NOTE: Set display=True 
word_labels = [int(k) for k in list(localized_words.keys())]

# Some Other Helpful Words related functions
# # Query a single word
from swtloc.configs import (IMAGE_ORIGINAL,
                            IMAGE_SWT_TRANSFORMED)
loc_word, swt_loc, orig_loc = swtImgObj.getWord(key=word_labels[8])

# # Iterating over all the words
# # Specifically useful for jupyter notebooks - Iterate over all
# # the words, at the same time visualizing the localizations
word_gen = swtImgObj.wordIterator()
loc_word, swt_loc, orig_loc = next(word_gen)

# # Generating a crop of a single word on any of the available
# # image codes
# # Crop on SWT Image
swtImgObj.saveCrop(save_path=respath, crop_of='words', crop_key=9, crop_on=IMAGE_SWT_TRANSFORMED, crop_type='bubble')
# # Crop on Original Image
swtImgObj.saveCrop(save_path=respath, crop_of='words', crop_key=9, crop_on=IMAGE_ORIGINAL, crop_type='bubble')