AN EFFECTIVE SELF-KNOWLEDGE DISTILLATION METHOD FOR WASTE CLASSIFICATION

Abstract

Garbage classification is a critical task in overcoming environmental pollution. People classify waste based on what they understand about the waste object rather than on its recyclable status, which leads to inaccurate classification in manual sorting. In addition, the direct processing of toxic waste can be physically dangerous for humans. Several machine learning and deep learning techniques have been proposed using standardized garbage classification datasets. However, these methods still have some disadvantages such as (i) the computational cost and memory of the proposed networks are large; (ii) they are difficult to run in real time; (iii) The training process is complicated because it is pre-trained on large data sets to avoid the phenomenon of "overfitting". In this paper, we propose a new approach based on self-knowledge distillation to overcome the above drawbacks. Experimental results have shown that our model achieves the best performance on the Trash Net dataset without having to pre-train the model. This proves the effectiveness of the self-knowledge distillation method brings to the problem of waste classification.