Waste Classifier (Optimized by OpenVINO & Neural Compute Stick 2)
- 0 Collaborators
Learn how to train a model and build a simple waste classifier that can be deployed on edge computing devices, optimized by Intel OpenVINO and Neural Compute Stick 2. The classifier can take an image of an item, classify it, and determine whether it belongs in a recycle, compost, or landfill bin. ...learn more
Project status: Published/In Market
Internet of Things, Artificial Intelligence
Intel Technologies
Movidius NCS,
OpenVINO
Overview / Usage
With increasing world population and volume of waste produced by humans, the need for Smart Waste Management solutions has never been more crucial than today.
This classifier was intended to assist humans in classifying their daily waste. From a use case standpoint, ideally it would be deployed in channels such as retail, smart cities, enterprises (eg. Intel, waste management companies), and etc. By adding another layer of computer supervision like this classifier, we can take proactive steps in keeping the world clean.
Methodology / Approach
Using TensorFlow Hub (machine learning library), the waste classifier was trained on a convolutional neural network known as Inception V3 (image classification). The dataset that was used in the training consist of images from Gary Thung and Mindy Yang's waste dataset (2527 images) and images found on the internet (4473 images).
The trained dataset is separated by 7 different classes, each class consisted of 1000 images. Cardboard (Mixed Recycling), Food (Compostable), Glass (Mixed Recycling), Metal (Mixed Recycling), Paper (Mixed Recycling), Plastic (Mixed Recycling), Trash (Landfill). Ideas such as feature engineering (targeting specific and more common waste items), maintaining consistent number of images in classes, and parameter tuning were used in the process of training this model
After training the model, it can successfully classify test images with reliable accuracy. The model can then be passed through OpenVINO's Model Optimizer and be used in the Inference Engine on one of its samples, Image Classification Python Sample. From the sample, the classifier can be specified to run on the Neural Compute Stick 2.
Technologies Used
- OpenVINO (Version 2019_R1.1)
- Neural Compute Stick 2
- Intel NUC Mini PC (Core i7) / Aaeon UP Squared (Atom)
- USB 2.0 Camera (take pictures)
- Ubuntu 16.04 LTS (Xenial Xerus)
- TensorFlow Image Retraining API
- Fatkun Batch Download Image (Google Chrome extension to quickly download images)