/AbInBevHackathon

Feature Extraction using Image Data

Primary LanguageJupyter Notebook

Problem Statement

How to gather insights on mislabelling of product and price compliance in stores across the globe ?

Insights gathered/ Final results:

  • The price tag of Brooklyn Indian Pale Ale and Brooklyn Oktoberfest were detected as interchanged in 9 different weeks
  • The price tag of Bud Light and Bud Light Platinum were detected as interchanged in 4 different weeks
  • With no mislabel Magic hat beer had maximum price mismatches
from ipywidgets import interact, interactive, fixed, widgets
from IPython.display import Image
import extract_template_matching
import matplotlib.pyplot as plt2
import cv2

import getting_ocr
import extract_max_contours
%matplotlib inline 
template_keys = ['Becks','BROOKLYN PILSNER','COORS LIGHT','LEFFE BLONDE','  MILLER LITE_6 Pk 12 Oz Glass','BLUE POINT TOASTED','BUD LIGHT','','','CORONA EXTRA','MAGIC HAT','SIERRA NEVADA PALE ALE','BROOKLYN INDIA PALE','BUD LIGHT PLATINUM','CORONA LIGHT','MILLER LITE_12 Pk 12 Oz Can','SIERRA NEVADA TORPEDO','BROOKLYN OKTOBERFEST','BUDWEISER','HOEGAARDEN','MILLER LITE_12 Pk 12 Oz Glass','TELLA ARTOIS']
sample_image = 'base_image.jpg'
loc = None

What was given??

Image(filename=sample_image, width=800,height=100)

jpeg

Challenge

  • Extracting the products in the shelf
  • Extracting the information from the price tag

Extracting the products from the shelf

  • Template Matching with Contour Detection
  • Features matching using deep net model, Inception (Scalable solution)

Extracting the information from the price tag

  • Tessaract OCR model for text extraction
  • ReInspect TensorBox model (Deep Net model for price tag detection and extraction)

Demo

def get_template_match(key='Becks', test_image=sample_image):
    template_image_path = 'sample_data/'+key+'.jpg'
    loc = extract_template_matching.get_template_matching(test_image,template_image_path)
interact(get_template_match,key=template_keys,__manual=loc)
<function __main__.get_template_match>

Image(filename='result.png', width=500,height=80)

png

price_tag = 'price_tag.jpg'
extracted_price_tag = 'extracted_price_tag.jpg'
Image(filename=price_tag, width=800,height=100)

jpeg

price_tag = 'extracted_price_tag.jpg'
Image(filename=price_tag, width=600,height=50)

jpeg

price_tag = 'price_tag_annotation.png'
Image(filename=price_tag, width=800,height=100)

png

import warnings
warnings.filterwarnings('ignore')
getting_ocr.get_tag_information([extracted_price_tag])
('Wrote', 2733, 'bytes to', 'jsons/extracted_price_tag.jpg.json')
---------------------------------------------
    Bounding Polygon:
{u'vertices': [{u'y': 7, u'x': 56}, {u'y': 7, u'x': 231}, {u'y': 58, u'x': 231}, {u'y': 58, u'x': 56}]}
    Text:
SIERRA NEVADA PALE ALE
13 49

Next Steps

  • Scalable solutions using Deep Learning methods like Inception and Reinspect model