On Building an Instagram Street Art Dataset and Detection Model

Last updated 201 days ago by Leonard Bogdonoff

python

What if you could pump all of the Instagram photos of Banksy's artwork into a program that could pinpoint where the next one's likely to be?

Well, we aren’t there quite yet, but there’s still some really cool stuff you can accomplish using image analysis and machine learning to better understand street art.

You can use machine learning models to detect whether an Instagram photo contains street art — even classify the type of street art. For example, you can make a classifier for stencil art, letterform, portrait murals, or mixed medium installations.

In this article, I will go over how to build a deep learning model using TensorFlow and Keras that accomplishes the task of generally detecting street art by using publicly available social media data on Instagram.

To my knowledge, there isn’t a publicly available dataset of street art or graffiti. But we’ll go over a few simple techniques for creating datasets from publicly available images on the Internet and social media — which will soon become indispensable tools in your machine learning toolkit. After reading this article, you’ll be able to leverage these methods to generate your own datasets for anything you need.

We’ll also learn how to build a TensorFlow model using Keras trained on our street art dataset. Then we will use this deep learning model to detect if new images contain street art photos.

Just pick an Instagram hashtag, grab some images, and train your deep learning model.

In the future, nearly everything will be photographed, and indirectly analyzable with machine learning. Learning how to train models to analyze this content yourself and understand the results is a superpower worth cultivating.

Overview of our Instagram street art dataset and model

Here’s a quick overview of our process:

  1. Build a street art deep learning image dataset using hashtag results for #streetart
  2. Use the images to build a deep learning model that will predict if images contain street art
  3. Clean the dataset and retrain the model for improved results

We’ll follow these three steps to build a real functioning model for classifying street art.The model here is based on the “Deep Residual Learning for Image Recognition” (2015) paper’s ResNet model and can be duplicated using other architectures.

You can view the finalized codebase in this Github repository. You can also open up the codebase (including the datasets I've collected) on FloydHub in a JupyterLab Workspace by clicking this button:

Read full Article