GarbageNet: A Unified Learning Framework for Robust Garbage Classification
Jianfei Yang,Zhaoyang Zeng,Kai Wang,Han Zou,Lihua Xie
DOI: https://doi.org/10.1109/tai.2021.3081055
2021-08-01
IEEE Transactions on Artificial Intelligence
Abstract:The recyclability of domestic waste plays a crucial role in the modern society, which helps reduce multiple types of pollution and brings economic effect. To achieve this goal, garbage classification is one of the most important steps during the recycling process. Prevailing deep learning techniques empower high-performance visual recognition models and can benefit the automation of garbage classification task. Nevertheless, there exist three challenges when directly leveraging deep recognition models for this task, referring to lack of sufficient data, high cost of category increment, and noisy data quality. In this article, we present a novel incremental learning framework, GarbageNet, to address the aforementioned challenges. First, weakly-supervised transfer learning guarantees the capacity of feature extractor. Second, for new categories of garbages, GarbageNet embeds them as anchors for reference and classifies the test samples by finding their nearest neighbors in the latent space. Third, an attentive mixup of training data is utilized for suppressing the negative effect of mislabeled data. We evaluate our method on real-world datasets, and the empirical results demonstrate that GarbageNet achieves the state-of-the-art performance with regard to accuracy, robustness, and extendability. The proposed method won the first place in the HUAWEI Cloud Garbage Classification Challenge in 2019. Impact StatementThis article contributes to a vision-based garbage recognition system for the recyclability of domestic waste. To the best of our knowledge, it is the first article that explores a realistic vision-based solution for the garbage recognition with a real-world dataset and a comprehensive benchmark on current state-of-the-art visual recognition models. The proposed method addresses several challenges and achieves state-of-the-art performance. We believe that this interdisciplinary research contributes to the AI for the environment, which helps promote environmental ethics, bring rotation economy, and relieve the pressure of consumption doctrine in smart city.