Modified Deep Learning Model for Efficient Recyclable Waste Classification: A Comparative Study of Convolutional Network Architectures

Abstract

Solid waste recycling management in Nigeria remains a challenge, necessitating efficient waste classification for environmental sustainability. This study proposes CiteWasteRN50, a modified ResNet50 Convolutional Neural Network (CNN), to classify waste into six groups. Key modifications include input image rescaling for enhanced feature extraction, three fully connected layers with dropout, and a SoftMax layer for probabilistic output. The model was trained and validated in MATLAB environment using the TrashNet dataset—75% for training over 7, 15, 30, and 45 epochs, and 25% for validation. CiteWasteRN50 was compared to eight other CNN models under identical conditions. At 15 epochs, CiteWasteRN50 achieved the highest classification accuracy of 98.09%, outperforming ResNet50, ResNet-18, DenseNet201, ResNet-101, EfficientNetB0, VggNet16, VggNet19, and InceptionV3 by 6.04%, 6.98%, 4.12%, 4.29%, 5.39%, 8.24%, 13.17%, and 5.08%, respectively. It also recorded the highest precision (0.9821), recall (0.9838), and F1-score (0.9819). Findings highlight CiteWasteRN50's strong accuracy and applicability for real-world waste classification.