A surprising amount of little-publicised academic effort is currently going into the problem of recognising brand logos in online images – specifically in the kind of chaotic, less-than-perfect imagery posted to social media in the millions every day.

Marked new interest in ‘visual listening’ and ‘virality detection’ indicates that logo diffusion is set to become a notable marketing indicator and analysis tool over the next five years; if it does, this has implications for the consumers who are providing the marketing medium for this relatively new field, and also for the companies and corporations which are interested in the social spread of their brands.

Specifically there is reasonable cause to believe that the company with a logo so distinctive that it can be recognised in even the most unfavourable lighting and general photographic conditions is likely to have an edge in data analysis – and that this factor might lead to new trends in logo design and placement.

Why companies want to track logos in social media

dhl-logo-in-social-media-photoThe value for a corporation in being able to track its logo across social media is, essentially, product placement. Since the overall nature of social media photos is celebratory or ‘positive’, and since the demographics of the core posting population are a marketing acme, significant work is going into new methods of delineating this data.

With over 1.8 billion photos posted every day, getting prominence for your brand in this gargantuan and influential stream is equivalent to a placement in a new Star Wars movie. In fact it’s significantly better, as neither the interest nor the intensity are ever likely to abate.

The obvious advantage of logo placement in social media is straightforward ‘ambient’ advertising; your brand appears in an image which you neither had to commission, take, nor transmit. You’re exposing the brand to potential customers when their pupils are fully dilated for other reasons (i.e. they’re looking at material that is meaningful and emotionally engaging for them to some or other extent – a level of ‘vulnerability’ practically impossible to generate in any other advertising context).

Logo recognition from social noise – a new SaaS proposition

Canadian social media company Sysomos has just acquired logo recognition startup gazeMetrix in order to leverage Deobrat Singh’s development work in recognising logos in images and videos:

sysomos-gaze-monitor

Tracking logo recognition in the console of Sysomos Gaze

Sysomos gazeMetrix is able to provide companies with bespoke feeds of where and when their brands appear in social media. GazeMetrix launched at the end of 2012 and first leveraged Instagram’s very flexible API to try and count the incidence of Starbucks coffee cups across posted images, and was surprised to find that the Starbucks logo was identified 10,000 times a day. “I didn’t believe it at first,” Singh said in 2013. “but we dug deeper into it and realized it was real—people were taking a lot of pictures of Starbucks mugs.”

Likewise mobile-based Ditto claims a 99% precision rate in identifying corporate logos in social imagery, with a SaaS provision, a social stream API and a cloud-based image recognition service for clients wishing to monitor particular social streams (nominally their own).

Dublin startup LogoGrab offers logo recognition technology developed by researchers from ETH, the Swiss Federal Institute of Technology, and claims to have developed ‘the most precise and scalable logo recognition technology’ on the market, offering ‘unprecedented rates of precision and accuracy’ and precise data on logo orientation, size, position and contextual relevance.

In January of this year LogoGrab CEO Luca Boschin declared the company to be ahead of the competition in being able to overcome the challenges in recognising logos, and gave the Nike logo as an example of a design that can confound image recognition capabilities: “It’s a really simple piece of visual information and lots of other things look like it. Traditional technologies in this realm can’t detect the Nike swoosh from other similarly shaped lines. Ours can.”

Applying Deep Neural Networks to the problem of logo recognition

‘An “inception meta-layer," as defined in GoogLeNet’’ - ‘DeepLogo: Hitting Logo Recognition with the Deep Neural Network Hammer’, Forrest N. Iandola et al.

‘An “inception meta-layer,” as defined in GoogLeNet’’ – ‘DeepLogo: Hitting Logo Recognition with the Deep Neural Network Hammer’, Forrest N. Iandola et al.

The problem with recognising a brand logo in an image is that it is ‘semi-analogue’ information; the logo is digital, but indistinct from its context. It is also unlikely to be presented in ideal conditions of lighting, perspective, proportion of entire image or general clarity. Additionally the logo may be only partially visible, a problem which some researchers are addressing specifically, as we’ll see shortly. Ultimately any shape can be represented by a hash or checksum which adds up adequate vector points to constitute a distinct configuration.

Colour adds another recognition hook, and colour pairs or groups seem likely to present an almost infallible ‘logo signature’ – which may mean some extra work at the design departments of Coca-Cola and Intel, among others, if accurate diffusion figures are to be obtained from social media photo streams.

The title of a recently published paper out of UoC Berkeley provides some litmus-style indication of how much money business is potentially prepared to put behind the science of logo-tracking. DeepLogo: Hitting Logo Recognition with the Deep Neural Network Hammer [PDF] presents research that turns a Deep Convolutional Neural Network (DCNN) loose on the much-used Flickrlogos-32 logo recognition dataset, along with the established ImageNet challenge dataset.

Forrest N. Iandola’s research group used Fast R-CNN (FRCN) to perform deep detection and localization of logos at faster and more accurate rates, it claims, than previous deep learning networks, succeeding also in simplifying the training process for the detection algorithms. The research touches on what is coming to be known as the ‘augmented reality’ aspect of logo detection, which involves logo recognition and localisation. In practice AR for logo recognition is currently leveraged by the LogoGrab app’s ability to recognise brand logos in photos and provide information about the related company, along with any special offers or other marketing material specific to the locale in which the photo was taken*.

Diagram of the Storefront Logo Recognition System’s ‘sliding window’ technique – ‘Fast and Robust Realtime Storefront Logo Recognition’, Frank Liu

Diagram of the Storefront Logo Recognition System’s ‘sliding window’ technique – ‘Fast and Robust Realtime Storefront Logo Recognition’, Frank Liu

Frank Liu of the Department of Electrical Engineering at Stanford University also produced research focusing specifically on the recognition of brand logos in storefront photographs, another likely scenario for the rich stream of daily social media uploads. Fast and Robust Realtime Storefront Logo Recognition [PDF] also developed an algorithm to specifically address the problem of occlusion – recognizing logos which are partially obscured by intervening objects (image right).

Logo redesigns for the image recognition age?

Since Google has a horse in this particular race, I thought it might be interesting to see if Google Images is able to recognise the first 20 of Complex’s Top 50 most iconic brand logos of all time, both at full and greatly reduced resolution. To make this ad hoc test I stripped out all metadata information from the photos that I challenged Google to identify, uploading first a full-size image and then the same photo at a maximum of 30 pixels width, as a logo might appear within a photo on someone’s shirt, or viewed from far away, or as co-branding on a coffee container. To help ensure I gave no clues to Google’s image upload process, I named the files randomly, kept all useful keywords (such as ‘logo’ in a folder name) out of the upload path and uploaded the lower-res images in a different browser over a different VPN address to the full-size images, with cookies freshly cleared first. Here are the results, which only feature the first twenty logos in the Complex article, due to lack of time – and bear in mind that these are the logos judged to be ‘most distinctive’, not ‘most successful’:

info-google-logos

Disney changed its logo 30 times between 1988 and 2015. Google is no richer or poorer for its logo rebrand this past summer. Companies often change their logos for irrational or whimsical reasons, so how much more might they be inclined to tweak them when there’s a rational one to hand? It’s even possible that vector checksums of new logos which have been designed for maximum resiliency across social media will become the subject of corporate patents.

For the end user, it’s probably just another day of being monetised in exchange for free services, with the EULA having anticipated all redress.


*Though it’s not a matter addressed in any of the research discussed here, a great deal of GPS information is uploaded by default with a social media photo, not only providing geographical metrics but, presumably, useful learning information for the active recognition algorithms, which are likely to improve their guesses when an ambiguous or indecipherable logo appears at coordinates where the same logo was previously recognised.