A Fusion of Three Custom-Tailored Deep Learning Architectures for Waste Classification

Population growth is exploding exponentially in this current world. Due to this escalating population and urbanization, the overall amount of waste is experiencing a rapid growth worldwide. Consequently, plenty of waste contributes to climate change, affecting our ecosystems and species. Fortunately, trash management would help alleviate some of these effects since a large quantity of trash is highly biodegradable and recyclable. However, classifying waste manually based on its contents is highly expensive and time-consuming. This is why the classification of wastes based on their contents is a critical criterion for ensuring cost-effective performance throughout recycling procedures. This paper proposes a hybrid deep-learning framework to classify waste into four categories: paper, glass, plastic, and organic. To address the lack of sufficient data, we used the albumentation function to augment the data. Later, in order to remove any duplicates that might have existed in the updated dataset, we also applied an image hashing technique that tackles the problem of overfitting. After preprocessing, we integrate three different models (2 EfficientNet models with noisy-student and imagenet and a custom convolutional neural network model) and provide prediction, and a heatmap of the eXplainable Artificial Intelligence (X-AI) generated images based on the test dataset to improve the trustworthiness of the inference. In comparison to various earlier state-of-the-art studies in the area of waste management, our technique performed substantially better, scoring at around 97% accuracy.