Development of a memory-efficient and computationally cost-effective CNN for smart waste classification
Iman Ameer Ahmad, Ahmed Mudheher Hasan, Amjad J. Humaidi · Journal of Engineering Research · 2025
Waste management is a critical limitation impacting environmental sustainability and public health, especially in developing countries where improper disposal leads to soil and water pollution. To control this growing problem, innovative solutions are needed that balance technological progress with environmental considerations. This study focuses on classifying solid waste materials, developing a detailed dataset of aluminum, cardboard, plastic, and glass samples collected from landfills and studio based. To process and classify these materials, we introduce an optimized Convolutional Neural Network (CNN) model using Global Average Pooling (GAP 2D) for better memory efficiency and speed-power-friendly, making it well-suited for real-time applications. Our custom CNN model, with only 495,620 parameters and a processing power of 1.27 GFLOPs, achieves an accuracy of 92.5%, balancing performance with efficiency. Our model stands out compared to leading models like VGG16, ResNet50, ResNet18, and YOLOv8. It outperforms VGG16 and ResNet50 in terms of test accuracy, precision, recall, and F1 score, all while being much more efficient in terms of computational cost, with only 1.27 GFLOPs compared to the much higher GFLOPs of VGG16 (30.72) and ResNet50 (7.77). Furthermore, our model has far fewer parameters, making it more resource-efficient: 495,620 parameters compared to 21 million in VGG16 and 49 million in ResNet50. While YOLOv8 achieves perfect scores for precision, recall, and F1 score, its high computational demands and large number of parameters make it less suitable for resource-limited environments. Overall, our custom CNN with GAP 2D offers an excellent solution for waste classification and other real-time applications, providing a strong balance between high accuracy and low computational cost.