Development of a CNN-Based Household Waste Image Classification Model for Automatic Identification

Abstrak

The problem of household waste in Indonesia remains a major challenge due to the low level of waste sorting and the lack of technology utilization in the waste identification process. This study aims to develop a Deep Learning-based inorganic waste identification system using the Convolutional Neural Network (CNN) method to improve the accuracy of automatic classification in the household environment. The research method used is an experiment with a quantitative approach. The dataset used consists of 92 images of inorganic waste covering the categories of plastic, paper, glass, and metal. The data went through a preprocessing stage in the form of image resizing to 224 × 224 pixels, normalization, and data augmentation, then divided into training and testing data with a ratio of 80:20. The CNN model was trained using variations in epoch parameters, batch size, and learning rate, then evaluated using metrics of accuracy, precision, recall, F1-score, and confusion matrix. The results showed that the training parameters had a significant effect on model performance. The optimal value was obtained at epoch 20 with 97% accuracy, a batch size of 8 with 97% accuracy, and a learning rate of 1×10⁻³ with 90% accuracy. The best parameter combination produced a model with 97% classification accuracy, along with balanced precision, recall, and F1-score values. These results demonstrate the effectiveness of the CNN method in automatically identifying inorganic waste types. The developed system has the potential to be applied in household environments to assist waste sorting, improve waste management efficiency, and support environmental conservation by raising public awareness of the importance of sustainable waste management. Although the model achieved 97% accuracy under a single 80:20 train-test split, its performance decreased substantially to 32.69% ± 6.93% under stratified 5-fold cross-validation, indicating potential overfitting and limited generalization.