A Comparison of MobileNetV2 and VGG16 Architectures with Transfer Learning for Multi-Class Image-Based Waste Classification

Effective waste management represents a global challenge with significant environmental and public health impacts. Despite existing waste classification systems achieving high accuracy rates, a critical research gap exists in determining optimal CNN architectures for real-world deployment constraints, particularly regarding computational efficiency versus classification accuracy trade-offs. We compared two Convolutional Neural Network (CNN) architectures MobileNetV2 and VGG16 for classifying ten types of waste using image-based analysis. Using transfer learning approach, both models were modified for waste classification tasks by adding custom layers to pre-trained models. The dataset contained 19,762 images balanced to 9,440 samples through under-sampling techniques and enhanced with data augmentation to increase variation. Results demonstrated that MobileNetV2 achieved 95.6% test accuracy with precision 0.93, recall 0.93, and F1-score 0.93, significantly outperforming VGG16's 89.13% accuracy with precision 0.91, recall 0.90, and F1-score 0.90. Beyond superior accuracy, MobileNetV2 also demonstrated higher computational efficiency with 350ms/step training time compared to VGG16's 700ms/step, and more consistent performance across all waste categories.