Enhanced YOLOv8 for Edge Computing: A Deep Learning Approach for Memory Optimization

Abstract

Enhanced YOLOv8 for Edge computing with a deep learning approach for memory optimization is proposed as an optimized Artificial Intelligence (AI) algorithm specifically designed for edge learning, which can efficiently enable Convolutional Neural Networks (CNN) deployment on Internet of Things (IoT) edge devices with less memory footprint and more rapid inference. This paper optimizes the YOLOv8 model architecture using the structured pruning method (Layer Pruning) to minimize non-contributing parameters. The proposed pruned model with loaded pre-trained model weights on the COCO dataset and after fine-tuning on the self-driving car dataset maintains high accuracy, with a negligible 0.011 decrease compared to the baseline, while reducing the model size by 33% and increasing detection speed to 196 Frame Per Second (FPS). These results are significant in making the proposed model well-suited for edge environments with limited memory capacity, especially for the dataset used allowing the vehicle to react rapidly to environmental changes for practical applications.