India has been facing a waste crisis in recent years. We generate around 62 million tonnes of waste a year, of which just 42 million tonnes are picked and a mere 12 million tonnes are treated, while the rest goes into a landfill. Many news outlets have reported a sharp rise in the amount of trash piling up in cities around the country, especially in urban centres. According to global forecasts, annual solid waste generation could exceed three billion metric tons by 2050, putting our environment and public health at risk.
Against this backdrop, an international team of researchers from the Ronin Institute in Montclair, USA; Asia University in Taichung, Taiwan; Symbiosis International University, Pune; Chandigarh University and others has developed a smart way to sort and organise waste more accurately and efficiently. By pooling expertise in computer engineering, medical research, data science, and other fields from around the world, these institutions aimed to address a universal problem - improving city waste management.
The researchers set out to create an automated 'smart' waste classification system, a tool that can tell whether a piece of garbage belongs in recycling, compost, or the trash, all by looking at a photo of it. They focused on six main garbage categories commonly found in city trash bins: Cardboard, Glass, Metal, Paper, Plastic, and General garbage.
They gathered over 2,000 images from a public collection (the Garbage Classification Dataset on Kaggle), which shows each of these waste types in different shapes and conditions. Next, they designed a computer model that extracts essential details from each image using a neural network called VGG16, a well-known tool in deep learning. After breaking the image down into features, like edges, shapes, and colors, those features were passed to an algorithm known as Random Forest, another popular machine-learning algorithm.
To improve the accuracy of the Random Forest algorithm in identifying the garbage, the team employed Cat Swarm Optimization (CSO). This unusual-sounding name refers to a mathematical method inspired by the behavior of cats (e.g., searching and tracking) to find the best values, or hyperparameters, for the Random Forest. Hyperparameters can be thought of as settings that determine how many decision trees are created and how deep those trees can grow, which in turn improves the prediction of the algorithm.
Convolutional Neural Networks (CNNs) like VGG16 examine images in layers. Each layer filters out different visual features, like edges, textures, patterns, etc., until the model has a detailed representation of what an image looks like. After passing through several layers, the image is transformed into a set of numbers that capture its most essential details.
Once an image has been converted into these numerical features, the Random Forest algorithm takes over. A Random Forest is made up of many smaller decision trees, each of which tries to classify the image into the correct category (like cardboard, glass, or metal). By averaging the decisions of all these trees, the Random Forest typically produces a stronger overall guess compared to just one tree. CSO helps the Random Forest find the best way to make decisions by adjusting parameters, so that misclassifications are minimized. The algorithm tries different combinations of these settings to see which one yields the highest accuracy.
By combining VGG16 for feature extraction, Random Forest for classification, and Cat Swarm Optimization for fine-tuning the Random Forest’s settings, the researchers ended up with a system that performed exceptionally well. They measured their success by checking the model’s accuracy (how many items are categorised correctly), recall, and the area under the curve (AUC), which measures the mean accuracy of predictions. Their final model was correct about 85% of the time, with an AUC of about 0.85, a very good score for this kind of automated sorting task.
The new approach to sorting trash uses a combination of modern AI tools, like deep learning (VGG16), machine learning (Random Forest), and an evolutionary optimisation method (Cat Swarm Optimization). Although there is some room for improvement, including integrating other kinds of sensor data and expanding the categories of waste it can sort, the research could make it much easier for city governments, waste management companies, and recycling programs to handle trash efficiently. Ultimately, an automated classification tool like this might help reduce the burden of waste on cities and offer a cleaner and greener environment.
This research article was written with the help of generative AI and edited by an editor at Research Matters.