A Classification Model for Small Waste Image Datasets With Knowledge Distillation and Input Masking
Ren Okubo, Masaomi Kimura · 2024
In recent years, there has been an increase in waste generation, leading to the problem of securing waste disposal sites. Therefore, various countries have advocated recycling to reduce waste generation. However, two issues remain in previous waste classification studies based on deep learning. First, improper labeling can be observed in a dataset, which makes waste classification models impractical. Second, achieving high accuracy with a small dataset using conventional waste classification models is challenging. To solve these issues, we create a dataset and introduce a knowledge distillation method to reduce the background impact and input masking approach to globally capture target object textures into a model combining pre-trained Vision Transformer and Convolutional Neural Network. We applied the proposed model to the dataset consisting of 3,217 waste images, and achieved an average accuracy of 95.30%. In addition, our proposed model resulted in a macro F1-score of 94.70%. These represent a 0.71% improvement over the average accuracy of the previous waste classification model and a 1.02% improvement over the macro F1-score of conventional model. The result demonstrates the potential of our approach to achieve a high generalization performance in waste classification with a small dataset.