Performance Evaluation of MobileNetV2 and DenseNet169 Using Transfer Learning for Inorganic Waste Image Classification

Abstract

Purpose: Manual waste sorting still has limitations because it requires considerable time and may cause errors in identifying waste categories. The development of Computer Vision and Deep Learning technology provides an opportunity to develop an automated waste image classification system. This study aims to compare the performance of two Convolutional Neural Network (CNN) architectures, namely MobileNetV2 and DenseNet169, for inorganic waste image classification using the Transfer Learning approach. Methods/Study design/approach: This study applies a Transfer Learning method using MobileNetV2 and DenseNet169 architectures. The dataset consists of inorganic waste images from two categories, namely cardboard and paper, obtained from the Kaggle dataset and primary documentation data. The research stages include dataset collection, image preprocessing, dataset splitting, model training, and performance evaluation using Accuracy, Precision, Recall, F1-Score, and Confusion Matrix. The dataset was divided into 80% training data, 10% validation data, and 10% testing data. Result/Findings: The experimental results show that MobileNetV2 achieved an Accuracy of 96.14%, Precision of 95.05%, Recall of 97.82%, and F1-Score of 96.42%. Meanwhile, DenseNet169 achieved an Accuracy of 96.14%, Precision of 95.70%, Recall of 97.09%, and F1-Score of 96.39%. Novelty/Originality/Value: This study provides an objective comparison between a lightweight CNN architecture (MobileNetV2) and a deeper architecture (DenseNet169) using the same experimental configuration. The results indicate that MobileNetV2 can achieve comparable classification performance with lower computational complexity, making it more suitable for implementation on resource-constrained devices.