HTE-YOLO: A Hybrid Transformer-Enhanced Deep Learning Framework for AI-Driven Real-Time Object Detection

The functions of real-time object detection on
mobile and edge devices are still considered a challenging task because of lack of computational power, memory management and energy conservation needs. In spite of the high performance of modern deep learning models, several of the state-of-the-art detectors have high computational costs and are not suitable to be deployed in lightweight settings. This paper presents a Hybrid Transformer-Enhanced YOLO (HTE-YOLO) model to be used in a mobile environment. It has an architecture that combines a lightweight YOLOv8 backbone and a Vision Transformer-based attention module to achieve a better grasp of the world around and higher representation of multi-scales features. Further, knowledge distillation strategy is used to enhance the generalization of the models without adding to the computation cost. The model is trained on a personal dataset of 8,000 images gathered in the real-world setting and tested on a part of the COCO dataset. Experimental performance shows better results of 99.2mAP0.5, 89.5mAP0.5:0.95 and an inference rate of 58 FPS using only 6.5M parameters. The suggested framework balances between accuracy and efficiency effectively, thus it is applicable to real-time mobile and edge.