Researchers From Trinity College Dublin Propose ‘WasteNet’: A Waste Classification Model Based On Convolutional Neural Networks

Climate change is one of the biggest challenges of our time. Waste recycling is a crucial process that helps mitigate this problem in two ways: it reduces the amount of utilized raw materials and the amount of waste that ends up in landfills. Our ability to effectively reduce and recycle waste has a lower environmental impact, and it is also less expensive. A waste classification solution should take place when waste is disposed of in a smart bin to avoid the contamination of potentially recyclable components. For instance, napkins, plastic bags, and coffee cups are among the most common contaminants in urban environments. An automatic waste classification system would solve this contamination problem while encouraging users’ interaction, as they should not worry about disposing of an item in the correct bin.

In this paper, a group of researchers from the Trinity College of Dublin proposed WasteNet, a deep learning model for waste classification based on the waste images that could be captured by a camera installed in the smart bin. The idea presented in this work is to deploy such a classifier at the edge to enable quick and intelligent decisions in smart bins without the need for cloud access.

WasteNet is trained through a hybrid transfer learning approach that consists of two stages. In the first stage of the training process, a network called DenseNet pre-trained on the ImageNet dataset is used as a fixed feature-extractor model. Specifically, the authors train this model on waste images by freezing the weights of its lower layers and only updating the weights of the top layers specialized for classification. This approach is one of the most used transfer learning methods for deep learning. Its purpose is to enable feature extraction for a new domain (waste images), starting from a source domain (the images contained in the ImageNet dataset).

Once the top layers have been accurately trained in the first stage, during the second stage of the training process, the other layers of the deep neural network are gradually unfrozen and fine-tuned, each with a different learning rate. This approach is called Discriminative Layer Training, and it aims at reducing the risk of catastrophic forgetting. Indeed, the higher-level layers of the network that have to learn domain-specific features are trained with a relatively high learning rate to quickly adapt their weights. On the other hand, the learning rate decreases while unfreezing the lowest layers, which learn more general image features (e.g., edges and corners).

WasteNet has been trained and evaluated through the TrashNet dataset that contains 2527 images belonging to six different classes: plastic, paper, glass, metal, cardboard, and others. The following figure shows how these images present different positions, orientations, and lighting conditions.

WasteNet achieves 97% accuracy on the 25% of the images of the dataset used to test it. Rarely, WasteNet wrongly predicts the class metal or the class plastic when, instead, the correct one should be glass. To further analyze the decisions made by WasteNet, the authors rely on a gradient-based eXplainable Artificial Intelligence (XAI) method. Such a method aims at highlighting the image areas that mostly contributed to the final prediction of the model. For instance, the following figure shows an image containing a glass bottle, wrongly predicted as a plastic one.

The model has focused on the neck and the base of the cylinder to correctly identify the bottle. Hence, the misclassification error probably happens since the dataset contains more plastic bottles than glass ones. WasteNet should thus be further trained on more images to mitigate these kinds of issues.

In conclusion, the current version of WasteNet is a good starting point for facing the automatic waste classification domain. However, as we have seen, the deep neural network presented in this paper should be further improved also to classify overlapping and moving items before being deployed in a real-world environment.

This Article Is Based On The Research Paper 'WasteNet: Waste Classification at the Edge for Smart Bins'. All Credit For This Research Goes To The Researchers of This Project. Check out the paper. 

Please Don't Forget To Join Our ML Subreddit