Enhancing YOLOv8n for Improved Small Object Detection on Custom Datasets

Abstract

Deep learning has achieved remarkable performance in object detection, with YOLO (You Only Look Once) standing out for its speed and accuracy. In this paper, we present an improved detection model based on the YOLOv8 architecture, evaluated on two large-scale datasets. Our method introduces a new detection scale (P2), enhancing small object detection by capturing finer features. Additional modifications include advanced upsampling, feature concatenation, and the integration of the C2f module into the model’s head, improving multi-scale fusion and overall accuracy. On a single-class dataset (9,215 images), our model achieves a mAP at 50 of 98.9% from scratch and 98.5% with pre-training, with precision and recall up to 97% and 96.2%, respectively. On a multi-class dataset with seven categories, it reaches a mAP at 50:95 of 78.1% with pre-training, and up to 94.5% precision and 89.6% recall. The model regularly surpasses YOLOv8n, YOLOv10n, and YOLOv11n across both datasets, exhibiting notable accuracy, robustness, and scalability, with a computational cost of 12.6 GFLOPs.