# Towards artificially intelligent recycling Improving image processing for waste classification

**Authors:** Youpeng Yu, Ryan Grammenos  
**arXiv:** [2108.06274](https://arxiv.org/abs/2108.06274)  
**Date:** 2021-08-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.06274v1/

## Key points

- Waste and recycling management industries are facing a significant challenge due to increasing global refuse
- There is a growing need for smart systems to monitor the environment and enhance recycling processes
- IBM's Wastenet project aims to improve recycling through artificial intelligence for waste classification
- Authors incorporate transfer learning and data augmentation techniques to improve accuracy of waste classification
- Convolutional neural network (CNN) used, with various training parameters tuned for optimal performance
- Model achieves impressive test accuracy of 91.21%
- Data augmentation techniques further enhance test accuracy to 95.40%
- Paper provides full details on model training and code is open-source for replication by other researchers
- Model successfully classifies recycling waste items in real-time using standard computer webcam
- Approach has potential for practical implementation in waste management systems

## Summary

The waste and recycling management industries are facing a significant challenge due to the increasing amount of global refuse. To address this issue, there is a growing need for smart systems that can monitor the environment and enhance recycling processes. IBM's Wastenet project aims to improve recycling by leveraging artificial intelligence for waste classification. In this study, the authors build upon the Wastenet project by incorporating transfer learning and data augmentation techniques to improve the accuracy of waste classification. They start with a convolutional neural network (CNN) and systematically select appropriate splitting ratios and tune various training parameters such as learning rate schedulers, layers freezing, batch sizes, and loss functions. The goal is to classify waste into different recycling types. To evaluate their approach, the authors compare and contrast results using 10-fold cross-validation. The model developed achieves an impressive test accuracy of 91.21%. Furthermore, they incorporate several data augmentation techniques including flipping, rotation, shearing, zooming, and brightness control. These augmentation techniques further enhance the test accuracy of the final model to 95.40%. What sets this paper apart from others in the field is its provision of full details regarding the training of the model. Additionally, the code for this work has been made open-source allowing other researchers to replicate and build upon their findings. One notable application of their model is its successful real-time classification of recycling waste items using a standard computer webcam which demonstrates its potential for practical implementation in waste management systems. Overall, this study presents an innovative approach to improving waste classification through artificial intelligence and data augmentation techniques with achieved high test accuracy suggesting that these methods have great potential for enhancing recycling processes and addressing challenges faced by waste management industries globally.