Scholarlink

<p>This study benchmarks four lightweight YOLO nano-variant object detection models, namely YOLOv5n, YOLOv8n, YOLOv11n, and YOLOv12n, using 701 thermal images of okra (<i>Abelmoschus esculentus</i>), comprising 355 adequately mature and 346 overripe samples. The models were evaluated using mean average precision (mAP@0.5−0.95), precision, recall, and inference latency across heterogeneous computing platforms, including GPU-accelerated parallel inference using an NVIDIA T4 GPU with TensorRT and sequential CPU execution using ONNX Runtime. This benchmarking addresses a critical gap in high-performance computing (HPC) deployment for non-RGB modalities, where real-time throughput exceeding 600&#xa0;FPS requires GPU-level parallelism and optimized operator fusion. Under a 10-epoch training protocol designed for resource-efficient deployment, YOLOv8n achieved the highest detection accuracy with a mean mAP of 66.3&#xa0;±&#xa0;0.3% and a GPU inference latency of 1.6&#xa0;ms, corresponding to a throughput exceeding 625&#xa0;FPS. YOLOv5n delivered comparable accuracy at 66.1&#xa0;±&#xa0;0.4% while exhibiting superior CPU inference performance of 31.1&#xa0;ms, making it well suited for edge deployment scenarios. Extended training with early stopping showed that attention-based architectures such as YOLOv11n (73.1%) and YOLOv12n (72.9%) achieve higher peak accuracy when longer training durations are permitted, whereas simpler architectures converge more rapidly. These findings provide practical guidance for selecting model architectures based on deployment constraints, particularly the trade-off between available training budgets and inference latency requirements in HPC-oriented agricultural inspection systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Benchmarking YOLO nano-architectures for real-time thermal imaging: application to okra maturity grading on heterogeneous computing platforms

  • 期刊论文

  • 发表日期：

    2026年1月24日

查看全文
  • Madasamy Raja Ganapathy,
  • Pathmanaban Pugazhendi,
  • Selvaraju Periasamy,
  • Basker Nagarajan

摘要

This study benchmarks four lightweight YOLO nano-variant object detection models, namely YOLOv5n, YOLOv8n, YOLOv11n, and YOLOv12n, using 701 thermal images of okra (Abelmoschus esculentus), comprising 355 adequately mature and 346 overripe samples. The models were evaluated using mean average precision (mAP@0.5−0.95), precision, recall, and inference latency across heterogeneous computing platforms, including GPU-accelerated parallel inference using an NVIDIA T4 GPU with TensorRT and sequential CPU execution using ONNX Runtime. This benchmarking addresses a critical gap in high-performance computing (HPC) deployment for non-RGB modalities, where real-time throughput exceeding 600 FPS requires GPU-level parallelism and optimized operator fusion. Under a 10-epoch training protocol designed for resource-efficient deployment, YOLOv8n achieved the highest detection accuracy with a mean mAP of 66.3 ± 0.3% and a GPU inference latency of 1.6 ms, corresponding to a throughput exceeding 625 FPS. YOLOv5n delivered comparable accuracy at 66.1 ± 0.4% while exhibiting superior CPU inference performance of 31.1 ms, making it well suited for edge deployment scenarios. Extended training with early stopping showed that attention-based architectures such as YOLOv11n (73.1%) and YOLOv12n (72.9%) achieve higher peak accuracy when longer training durations are permitted, whereas simpler architectures converge more rapidly. These findings provide practical guidance for selecting model architectures based on deployment constraints, particularly the trade-off between available training budgets and inference latency requirements in HPC-oriented agricultural inspection systems.