Architectural Evolution and Deployment Challenges in YOLOV۸-YOLOV۱۲: A Narrative Review

YOLO models constitute a widely accepted sequence of real-time object detectors, with recent iterations exhibiting advancements that go far beyond mere improvements in detection accuracy. This review attempts to evaluate YOLOV۸ through YOLOv۱۲ based on architectural innovation, training process, inference operations, efficiency, maturity of evidence, and practical deployment feasibility. YOLOV۸ is considered a practical modem baseline due to extensive ecosystem and ease of deployment support. YOLOV۹ is analyzed as a training-focused model with a focus on information transfer using programmable gradient information (PGI) and generalized efficient layer aggregation network (GELAN). YOLOv۱۰ is reviewed as an end-to-end detection design minimizing NMS reliance. YOLOv۱۱ is critically analyzed as primarily documented by a technical report and ecosystem support, emphasizing scalable aggregation and multi-task adaptation. YOLOv۱۲ is reviewed as a relatively recent attention-based real-time object detection design with the focus on area attention, R-ELAN, and efficient attention computation. Contrary to conventional benchmark-based rankings, this article attempts to provide a bottleneck-oriented interpretation of YOLOV۸-YOLOv۱۲, in which each successive iteration targets a particular bottleneck stage in the detection pipeline. This article concludes that practical choice of YOLO variant should take into account not only mAP but also runtime, hardware, memory accesses, post-processing, software maturity, and reproducibility.

کلیدواژه ها:

YOLO ، object detection ، real-time detection ، edge deployment ، attention mechanisms