EVALUATING DEEP LEARNING MODELS FOR REAL-TIME WASTE CLASSIFICATION IN SMART IOT ENVIRONMENT

Abstract

Automatic municipal solid waste management systems are integral to  every smart city worldwide. They help to separate wastes into different categories for further recycling or effective disposal. This way, waste authorities could mitigate the effect of rapid urbanization, population growth, and the escalating consumption patterns associated with modern living. Deep learning models could play a critical role in the identification and classification of these wastes into their respective categories. Therefore, this study evaluates the performance of three deep learning models: MobileNet, InceptionV3, and VGG16 for waste classification. The evaluation was done under two separate model configurations while their classification accuracy, execution time, precision, recall, and F1-Score were computed across a range of 10 to 100 epochs. MobileNet consistently demonstrated the highest classification accuracy, reaching approximately 90% at 100 epochs, while also maintaining the shortest execution time, starting at 2.13 minutes for 10 epochs and increasing to about 14.34 minutes for 100 epochs. InceptionV3 exhibited a balanced performance, achieving around 83% accuracy at 100 epochs with execution times ranging from 3.57 minutes to 49.42 minutes. VGG16, although started with the lowest accuracy, improved significantly to about 88% at 100 epochs, but at the cost of the longest execution time, starting at 9.45 minutes and rising to 68.72 minutes. The results indicate that MobileNet is the most efficient model for applications requiring both high accuracy and low computational cost, while InceptionV3 and VGG16 are suitable for scenarios where accuracy is prioritized over execution time. This comparative analysis provides valuable insights for selecting appropriate deep-learning models based on specific task requirements and resource constraints.