Analysis the improvements of YOLOv5 algorithms: NRT-YOLO, MR-YOLO and YPH-YOLOv5

Abstract

Computer Vision (CV) is a fundamental aspect of artificial intelligence, with applications spanning multiple domains. The YOLO (You Only Look Once) algorithm has significantly contributed to real-time object recognition in CV. This paper explores the evolution of the YOLO algorithm, focusing on the improvements brought by three specialized variants: NRT-YOLO, MR-YOLO, and TPH-YOLOv5. NRT-YOLO addresses the challenge by introducing the C3NRT module, enhancing precision while maintaining low complexity. MR-YOLO optimizes YOLOv5 for industrial quality control, improving speed and accuracy. TPH-YOLOv5 enhances object detection in drone-captured images by introducing additional prediction heads and transformer modules. Despite these advancements, YOLO algorithms have limitations, including difficulty in detecting small objects and issues in complex scenes. Nevertheless, the research sheds light on the continuous evolution of YOLO algorithms, offering insights into real-time object detection, with applications in manufacturing, healthcare, transportation, and more. The significance of this research lies in showcasing the adaptability and potential of the YOLO framework. As YOLO continues to evolve, it promises to revolutionize various industries and applications, providing robust, real-time object detection solutions.