PENGEMBANGAN SISTEM KLASIFIKASI SAMPAH UNTUK DAUR ULANG OTOMATIS BERBASIS CITRA MENGGUNAKAN MODEL CNN
Faqih Rifaldy Faqih Rifaldy, M Irsan Prayoga M Irsan Prayoga, Lailan Sofinah Harahap · Melek IT Information Technology Journal · 2025
One of the most significant environmental issues that needs to be carefully considered is waste, especially when it comes to recycling. The goal of this study is to create an automated waste classification system that uses images to aid the recycling process. Using a large dataset that includes over 30 categories, a Convolutional Neural Network (CNN) model is used to categorize various waste materials, including paper, metal, glass, and plastic. To improve the generalization of the model, the dataset is preprocessed using several methods, including data augmentation. Rotation, flipping, and zoom adjustment are examples of augmentation techniques that generate more data. Visual elements of the waste images are extracted using a straightforward CNN model architecture that includes multiple convolutional and pooling layers. The Adam optimizer is used to train the model for ten epochs with a learning rate of 0.001. The overall accuracy of the model on the test dataset is 66%, according to the evaluation results, with varying precision and recall by category. According to the confusion matrix analysis, the model performs well on the plastic and glass categories, but has difficulty distinguishing between classes that have similar visual characteristics, such as metal and paper. The Histogram of Oriented Gradients (HOG) feature is also used in the Support Vector Machine (SVM) model as a basis. Deep learning-based methods are preferred because the SVM model only achieves 55% accuracy, lower than CNN. Furthermore, a comprehensive assessment is carried out using f1-score analysis and prediction visualization for each class. This research is an initial step towards creating an image processing-based recycling automation system. With the accuracy achieved, this model can be a basis or reference point for further research. In the future, data on underperforming classes will be added, transfer learning models such as EfficientNet will be used, and the model will be integrated with real-time systems for beneficial applications. It is hoped that the findings of this study will provide a basis for improving the effectiveness of waste management and assisting environmental conservation initiatives.