Ingeniería Electrónica

Análisis comparativo de modelos supervisados y semisupervisados para la clasificación de residuos mediante visión por computador

Automatic waste classification using deep learning techniques represents an alternative for supporting solid waste separation and management processes. However, supervised learning models depend on labeled datasets, whose construction may require considerable amounts of time and resources. This work compares ordinary and post-consumer waste classification models based on convolutional neural networks using supervised and semi-supervised learning approaches. For supervised training, a hyperparameter tuning process is conducted and different convolutional architectures are evaluated, leading to the selection of ResNet-18 and MobileNetV2 for subsequent experiments. For the semi-supervised approach, FixMatch is implemented using an 80/20 split between unlabeled and labeled data, and two modifications to the pseudo-labeling process are evaluated: a dynamic threshold scheme and the incorporation of Mean Teacher. The results show that the best supervised model achieves an accuracy of 96,82%, while the dynamic threshold scheme with ResNet-18 obtains the best semi-supervised result, reaching an accuracy of 94,68% and a relative difference of approximately 2,21% compared with the best supervised model, while using only 20% of the labeled data employed by the latter. Furthermore, MobileNetV2 presents shorter training times than ResNet-18 across the evaluated configurations, with reductions of approximately 40% to 45% in some semi-supervised scenarios. Overall, the results show that the selection of the architecture and the pseudo-labeling process leads to different trade-offs among accuracy, amount of labeled data, and training time, demonstrating that semi-supervised learning represents an alternative for reducing dependence on manual labeling while maintaining classification performance close to that obtained through supervised learning.

Palabras clave

Tesauro de su biblioteca