sid597/vadapav

My implementation to Build and train a system to detect if a given image is of a Vada-pav or not!

Vada Pav

My implementation to Build and train a system to detect if a given image is of a Vada-pav or not!

Project started : 29 dec 2019

Steps I took to scrape data :

• Searched for dataset in kaggle and google but did not find there
• Checked out the following platforms for images but did not find good results :
  • Instagram : search #Vadapav but results were not very meaningful
  • Facebook : same as Instagram did not find good results
  • ShutterStock : Paid service so did not use it
• Checked Google and Bing for image searches and there found much better results

How did I get images from google and Bing ?

I did it in 2 parts :

• Get urls of all the pics in search result
• Go to the links and download them

1. To get the list of urls I went to the image search results and scrolled till the results which came were revelent. I did this manully could have also done through selenium by scrolling the document but I took this approach Then in the developer consolse I wrote script to extract all these urls and download them in a text file the script is in get_urls.txt and the urls from Bing and Google can be found in bingUrls.txt and googleUrls.

2. Wrote scripts to download images from bing and google and saved them in bing_images and google_images. The scripts are in [download_from_bing.py] (./download_from_bing.py) and download_from_google.py

** After I did this proces I found out an extension called Fatkun Batch Download Image which makes this process very easy Searched for plugins before also but don't know how this one skipped from me

Found about 800 images combined from both google and bing. Then I manually checked and discarded the images now have ~500 meaningful images overall. Kept 80 images for test and ~420 for training.

Images obtained were of different sizes so wrote a script to make the resolution smalland same for all which can be found inside resize.py. Used PIL library to resize the images in 300x300 resolution docker run -it tensorflow/tensorflow:1.1.0 bash

Augumentation

Augumented the images to have rotated and flipped images scripts for them are in transform.py Now have a total of ~2000 images for training

For model training

For training I am going to use retrain.py provided by google tenserflow Steps I am going to follow :

• Create a folder named 'images' and Inside it will have 2 folders one for trainigs images and the other with test images

• Download and retrain the Inception V3 net by google

• Run the following command

  python retrain.py
  --bottleneck_dir=bottlenecks
  --model_dir=inception
  --summaries_dir=training_summaries/long
  --output_graph=retrained_graph.pb
  --output_labels=retrained_labels.txt
  --image_dir=images

• This will generate my own image classifier for Vada pav.