Leveraging on the Transfer Learning with ResNet-50 for Efficient Classification of Waste Categories

Abstract

The rapid urbanization and population expansion in Nigeria have resulted in a substantial rise in solid waste production, posing considerable issues for efficient waste management and environmental sustainability. This study addresses these difficulties by creating an automated trash classification system utilizing deep learning techniques to improve the efficiency and precision of waste classification. The ResNet-50 convolutional neural network was utilized via transfer learning to train a model using a meticulously assembled dataset of 2,527 waste images categorized into six types: cardboard, glass, metal, paper, plastic and trash. Data augmentation techniques, including random rotation, flipping, and zooming, were employed during preprocessing to enhance model generalization and resilience. The training procedure utilized a two-stage approach: first, the base ResNet-50 layers were frozen to leverage pre-acquired general image data, and subsequently, the top layers were fine-tuned to accommodate waste-specific attributes. Evaluation criteria such asccuracy, precision, recall, and F1-score showed consistently strong performance, with validation accuracy maintained at 97.8%. The confusion matrix demonstrated robust classification performance. The findings highlight the model's capability to efficiently automate waste classification, hence diminishing reliance on labor-intensive manual sorting commonly found in Nigeria. This study advances the overarching objective of sustainable urban trash management by offering a scalable, precise, and economical classification methodology.