| Popis: |
As population grows, there is an exponential increase in garbage generated. This garbage contains a high percentage of plastic that is recyclable. Therefore, it is necessary to segregate different types of garbage. To minimize the impact of incorrect garbage separation, we propose using an automated system to correct the waste classification. We studied in-depth learning techniques by using comparison-based experiments. A ResNet-152 and ResNet-50 were investigated to obtain the best-fit classification model for garbage classification. The model was trained on two datasets, TrashNet and local datasets which contains 5,326 images of four garbage categories (metal, pet, plastic, and trash). The accuracies for the pre-training hyperparameters with classification time on average were 98.99% with 0.27 s and 98.81% with 0.72 s for ResNet-152 and ResNet-50, respectively. The experiment testing results on both ResNet-152 and ResNet-50 yielded outperformed accuracy respectively by 95.72% and 94.15%. Using computer vision and IoT-based system, we implemented the selected model having more steady learning rate, ResNet-50, on the IoT devices to simulate the automated garbage segregation machine. The Raspberry Pi camera connected on a microcontroller captures one by one image of the garbage when the motion sensor triggers it. The captured image is then sent to the classification model by returning a predicted garbage class displayed on the output screen. Our future works based on the output generating a garbage category. The final classification model would be integrated with a fully automated sorting machine by automatically moving to its respective bin using a motorized sliding tray. |
