Identifying crowdsourced images to incentivize recycling with AI


Founded: 2018

AI problem

Multi-label image classification



Read about how our user Bower, the app incentivizing its users to recycle more, built an AI model on the Peltarion platform to enable their expansion throughout Europe.

02/ Incentivizing recyling

Bower (formerly PantaPå) is a Swedish-based company working towards the vision of a better recycling world. The company has developed an app incentivizing the recycling of household waste.

Every month, Bower’s users recycle around 1.5 million packages, and all of them are recycled at any of their 30k+ recycling stations. 

The company’s AI journey began with its expansion plans to additional markets in Europe. They needed a way to automate the approval process of new images of recycling stations that were being uploaded by users, as this would otherwise become too big of a task to manage manually.

03/ Automating the workflow

Before Bower started to look towards AI to enable them in their work, a human operator had to manually go through each image of recycling stations being sent to them by users, awaiting approval. This was a tedious job, consisting of a few hundred images a day. 

The uploaded images had to be checked so that these were in fact images of a legit recycling station, i.e. so that it was not just an image of a photo of a screen portraying one of these stations, or something entirely different, like a close-up shot of a garbage bag. Once confirmed that the image actually portrayed a recycling station, they also had to make sure it was in fact one of the approved recycling stations (there are several). 

04/ The opportunity for deep learning

This problem was a perfect case to use image classification for. The input is an image and the output is a category, in this case, the type of recycling station. Additionally, a separate model can be used to filter out images that are not recycling stations at all, as a binary image classification task (yes/no).

A convolutional neural network is trained using examples of the various recycling stations (photos) so that eventually it is able to understand the content of the image and predict with high enough accuracy the correct type of recycling station (if any) present in the picture.

05/ Data requirements

Because it is possible to use a model with pre-trained weights, the number of examples to get good performances with this model is relatively limited. One hundred independent examples for each category is probably enough to achieve good accuracy. For some categories that had more variance (i.e. items of that category could look pretty different from each other) we could have used additional data (between 50 and 100 % more, depending on the severity).

06/ The results

Thanks to the modul built, Bower managed to build a solution that cuts down the time we spend on reviewing these stations by 75% and semi-automating the review process, meaning they are still in control in certain cases. Additionally, because this is now an area where less manual work is needed, they can instead focus on more important tasks and enabling their expansion across Europe.

07/ What else can be done