Clients brief

Automate waste sorting process with an AI-based program and computer vision.Identify plastic bottles among other waste with an accuracy of over 50%.Classify identified plastic bottles.

What does the app do

Ralabs team created a convenient product that helps to detect plastic bottles amongthe other waste. The application identifies five types of bottles: transparent, milk,chemics, green, and brown. It can be connected to the manipulator that takes selectedwaste from the conveyor. Our application help to automate a waste sorting process and reduce expensesfor a recycling company.
What does the app do - The Waste Detector

How we did it

During the testing period, we were trying different detector models to find thebest solution. We used Faster Regional Convolutional Neural Network (FasterRCNN),You Only Look Once (YOLO), Region-based Fully Convolutional Network (R-FCN),and Single Shot Detector (MobileNet SSD). The task was not just to train a model todetect plastic bottles and sort them. Our main goal was to make a level of recognitionaccuracy higher than 50%. We started our model training from 250 photos of the waste. In a process, we increasethis number to four thousand photos that made a dataset more accurate. Our teamfound out that FasterRCNN is less demanding to the data and more correct than YOLO,MobileNet SSD or R-FCN, but slower than other models. YOLO the fastest system,but not accurate enough. Thus, we choose R-FCN that quiсker than FasterRCNN andmore trustworthy than YOLO. As a result, the current average accuracy is 87%.
How we did it - The Waste Detector

More related cases.

case-banking-cr2-logo
The unique omnichannel banking platform for all mobile technologies.
View Case Study

Mobile Cryptocurrencies Wallet

The interactive mobile application which allows to buy and store over two dozen types of cryptocurrency.
View Case Study