Benchmarking Modern CNN Architectures for Transfer Learning-Based Recyclable Waste Classification

Abstract

Accurate waste sorting is essential for improving recycling efficiency 
and supporting sustainable waste management. However, manual 
sorting remains slow, labor-intensive, and vulnerable to human error, 
particularly when recyclable materials have similar visual 
appearances. Deep learning offers a promising solution by enabling 
automatic image-based waste classification. This study benchmarks 
six modern Convolutional Neural Network (CNN) architectures, 
namely EfficientNetB0, ResNet50, InceptionV3, MobileNetV2, 
MobileNetV3, and ShuffleNet, for classifying recyclable waste using 
a transfer learning approach. A total of 4,090 images from four 
categories, glass, cans, paper, and plastic, were used and divided into 
training, validation, and testing sets with an 80:10:10 ratio. All models 
were trained under the same configuration, including 20 epochs, a 
batch size of 32, the Adam optimizer, and a learning rate of 0.0001. 
Performance was evaluated using accuracy, precision, recall, and F1
score. The results show that EfficientNetB0 achieved the best 
performance, with 98.29% accuracy, 98.25% precision, 98.25% 
recall, and 98.25% F1-score, followed by ResNet50 with 97.80% 
accuracy. In contrast, ShuffleNet produced the lowest accuracy of 
72.37%. These findings indicate that EfficientNetB0 is the most 
effective architecture in this study and has strong potential for 
automated recyclable waste classification systems.