Application of improved YOLOv5 model on object detection across multiple datasets

This paper presents a series of enhancements made to the YOLOv5 model, which belongs to the well-established YOLO (You Only Look Once) series of object detection models. The proposed modifications yield a significantly advanced object detector exhibiting exceptional performance across diverse datasets. The primary focus of our improvements lies in the replacement of select convolutions within the model using an enhanced reparameterization technique tailored for convolutional models. In conjunction with other effective enhancement strategies, the augmented YOLOv5n model achieves a mean average precision (mAP) of 77.8 on the VOC2007 dataset, showcasing an impressive 18% performance gain over the original model (version 6.0). This notable improvement positions YOLOv5n ahead of the state-of-the-art YOLOv8 model, while concurrently attaining further enhancements in frames per second (FPS) compared to its predecessor. A comprehensive set of experimental results substantiates the efficacy of our approach towards enhancing the YOLOv5 model, rendering it more amenable to the requirements posed by various application domains within the field of object detection.