Optimized YOLOv8 for multi-scale object detection

Object detection is considered one of the main tasks in computer vision and finds wide application in various fields, including medical imaging, face detection, object recognition, and many others. With the advent of AI, most contemporary object detection approaches rely on CNN methods such as Faster R-CNN and YOLO. YOLOv8 is one of the most renowned object detection algorithms, acclaimed for its ability to quickly and accurately detect objects of varying sizes, from small to large. YOLOv8 offers five variants, the smallest comprising 225 layers. Utilizing YOLOv8 for specific object sizes and resource-constrained applications may entail computational costs. In this paper, we introduce six modified versions of YOLOv8 tailored for different object sizes: small, medium, large, small–medium, medium–large, and small–large. These proposed models are evaluated based on computational cost, energy usage, and mAP-50, demonstrating reduced computational overhead compared to the original while maintaining accuracy.