Real-time deep neural network-based waste detection and classification using a camera sensor

Abstract

Waste generation is a growing environmental concern, with manual sorting methods often being inefficient and error-prone, particularly under varying lighting and environmental conditions. In Indonesia, waste is typically categorized into organic and nonorganic, yet existing automated classification systems lack real-time capabilities and robustness in dynamic settings. This study proposes a novel real-time waste detection and classification system using a deep neural network, implemented on the Jetson Nano platform with a camera sensor. The system utilizes the ResNet-18 convolutional neural network architecture and is developed using Python. It is designed to distinguish between organic and nonorganic waste in real-time. Training was conducted over 30 epochs, and the system was tested under various lighting conditions—morning, daytime, afternoon, and nighttime. Results show high accuracy: 95.24% in the morning, 95.24% during the day, 90.45% in the afternoon, and 86.90% at night, with an average accuracy of 91.96%. Performance was influenced by factors such as lighting intensity, distance, waste position, changes in organic waste, and occlusion by plastic. The proposed system offers a significant improvement over traditional and existing methods by enabling accurate, real-time waste classification under diverse conditions, contributing to more efficient and intelligent waste management.