Tech mahindra
Tech mahindra

The Rise of Image Analytics: AI in Social Media

Posted by: Niranjan On June 07, 2018 03:17 PM facebook linked in twitter

Computer vision has been a fundamental problem that the AI community has grappled with from the start of the AI age from the mid 1950’s and through its decades of summers and winters. While there are many tasks in vision such as recognition, motion analysis, scene reconstruction and image restoration, it is recognition that is the most fundamental.

From the late 1960s till relatively recently, though the theory was sound and the uses plentiful, the main places where this was noticeable was inside laboratories and in difficult to implement OCR systems. The initial work done by Hubel and Weisel on animal visual cortices and processing made its way into AI systems inspiring several generations of image recognition and classification methodologies, primarily around convolutional neural networks (CNN – while there are new and old contenders to try and displace CNN like Deep Forest, cascade classifiers etc. none have displaced CNN till date ). These in turn were utilised for various recognition sub-tasks such as content-based retrieval, OCR, facial recognition and shape recognition.

The breakthrough for image recognition came in 2011-12. CNN used for image classification of an MNSIT database showed error rates as low as 0.23% and was very fast. Since then, teams worldwide have been participating in an Imagenet Large Scale Visual recognition project (ILSVRC) to compete in increasingly difficult visual recognition tasks and they have been getting better with every succeeding year! The current state is so advanced that there is specific hardware called Vision Processing Units (VPUs) to help run algorithms like CNN and SIFT (scale variant feature transform)

The main reason for this explosion in Visual Recognition which falls under a categorisation called narrow AI, has been

  1. Massive open datasets that can serve as training inputs,
  2. Open source frameworks and massive worldwide collaboration
  3. API services, hitherto unknown, to boost solution creators for end user scenarios

This industry in its entirety is poised to change humanity’s worldview with use cases ranging from inputs for autonomous vehicles, to vision systems for domestic and industrial robots, facial recognition for law enforcement and automated cross platform metadata tagging amongst others. One place where it has found a natural home has been in the social listening and analytics ecosystem. It is a near perfect fit due to:

  1. The increasingly visual nature of social networks and information dissemination (~ 3.2 billion images are shared online every day - Meeker)
  2. Underreporting of non-text based sources in the insights space
  3. A need for visual understanding of logos, product placement and how products are actually being used in real life

Since 2017, a few mainstream social-insights providers have already ventured into the space – yes, powered by CNN. If you are a brand marketer, who is looking to get into this space, some of the areas you could start working exploring right away are:

  1. Inspiration from points of product consumption
  2. Sponsorship ROI with logo recognition (offline logos captured in the multitude of images can now be parsed and analysed)
  3. Integrate data-driven visual imagery into brand storytelling
  4. Including visual imagery to gauge virality (will help predict events with a completely utilised and rich data source)

The top things that you need to keep in mind are

  1. Whether the visual recognition system can seamlessly integrate into the text analytics offerings in place (integrated views)
  2. How efficient is the system. Typically, systems that are trained for specific scenarios tend to fare better than generic systems. So, testing and pilots, especially in the non-first world contexts are important
  3. How deep is the visual content understanding – it should ideally encompass logos, objects, scenes, facial characteristics and actions
  4. How quickly and easily can the visual search system be setup?
  5. Most importantly, how do you identify and integrate the implications of new data sources and their insights into the social analytics sphere, which is a mostly defined process in most large enterprises.

Sources

https://www.allaboutcircuits.com/news/3-applications-for-ai-image-recognition/

https://www.wired.com/story/researcher-fooled-a-google-ai-into-thinking-a-rifle-was-a-helicopter/

https://www.brandwatch.com/blog/image-analytics-future-social-listening/

https://www.scraawl.com/product/2017/09/20/image-recognition-analytics-digital-marketing/

http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6248110

http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture5.pdf

https://news.ycombinator.com/item?id=13773127

https://d3v6gwebjc7bm7.cloudfront.net/event/15/98/94/9/rt/1/documents /resourceList1519239358183/imageanalyticswebinar1519239372747.pdf

https://en.wikipedia.org/wiki/ImageNet#ImageNet_Challenge

About Author

Niranjan, Digital Solutions

Niranjan (NJ) is an experienced, commercially-driven digital marketing practitioner and consultant with more than 12 years of experience. He has worked with organizations ranging start-ups to Fortune 10 organizations in the digital marketing office. He is a consultant for clients across the US, UK, Asia Pacific, Middle East and Japan. Data, Marketing, and Technology are his USP and has vast experience with tools such as Salesforce Marketing Cloud, Adobe Marketing Cloud, and a plethora of "big data" and digital marketing/social media tools.

Tags: Connected Platforms & Solutions
 
(*) symbol is mandatory field
Name:
* Email Address:

Comments

(*) symbol is mandatory field

Post a Comment

* First Name:
* Email Address:
URL:
Comments:
 
Image Code
* Enter Image Code



Contact Us

Archives

For further information please write to connect@techmahindra.com

For further information please write to connect@techmahindra.com