Scholarlink

<p>This paper presents a novel edge-deployable fabric hue recognition system designed for real-time industrial applications. Our proposed framework integrates a YOLO-based object detection model with an enhanced lightweight convolutional neural network (CNN). This cascaded architecture efficiently localizes fabric regions using YOLO, then performs hue classification on the detected areas. By processing only the region of interest, the system minimizes redundant computations and sustains high throughput for multi-object scenarios. The classification model, which leverages depthwise separable convolutions and operates in the HSV color space, with the extracted features subsequently classified by a radial basis function (RBF) kernel support vector machine (SVM), is optimized for resource-constrained edge devices. The choice of HSV over RGB is empirically validated: Our CNN+SVM pipeline achieves 98.02% accuracy in the HSV space compared to 96.18% in RGB, while also requiring fewer SVM support vectors (79 vs. 109), confirming that HSV provides cleaner inter-class separation for fine-grained hue discrimination. To validate the system’s robustness under realistic conditions, we evaluated its performance on a custom-built conveyor belt testbed with controlled illumination. The system demonstrates strong resilience to ambient light variations and achieves a 98.02% accuracy on a real-time video test set. Edge deployment on an NVIDIA Jetson AGX Xavier with TensorRT acceleration yields 558.87 FPS with an average per-frame latency of 1.79 ms, confirming the system’s suitability for real-time on-site deployment. This work provides a low-cost, reliable, and computationally efficient solution, making it highly suitable for rapid, on-site fabric hue inspection on production lines.</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

Real-time fabric hue classification in industrial environments using YOLO-assisted lightweight CNN and RBF-SVM

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Feng-Cheng Lin,
  • Cheng-Xun Yu,
  • Chih-Cheng Chen

摘要

This paper presents a novel edge-deployable fabric hue recognition system designed for real-time industrial applications. Our proposed framework integrates a YOLO-based object detection model with an enhanced lightweight convolutional neural network (CNN). This cascaded architecture efficiently localizes fabric regions using YOLO, then performs hue classification on the detected areas. By processing only the region of interest, the system minimizes redundant computations and sustains high throughput for multi-object scenarios. The classification model, which leverages depthwise separable convolutions and operates in the HSV color space, with the extracted features subsequently classified by a radial basis function (RBF) kernel support vector machine (SVM), is optimized for resource-constrained edge devices. The choice of HSV over RGB is empirically validated: Our CNN+SVM pipeline achieves 98.02% accuracy in the HSV space compared to 96.18% in RGB, while also requiring fewer SVM support vectors (79 vs. 109), confirming that HSV provides cleaner inter-class separation for fine-grained hue discrimination. To validate the system’s robustness under realistic conditions, we evaluated its performance on a custom-built conveyor belt testbed with controlled illumination. The system demonstrates strong resilience to ambient light variations and achieves a 98.02% accuracy on a real-time video test set. Edge deployment on an NVIDIA Jetson AGX Xavier with TensorRT acceleration yields 558.87 FPS with an average per-frame latency of 1.79 ms, confirming the system’s suitability for real-time on-site deployment. This work provides a low-cost, reliable, and computationally efficient solution, making it highly suitable for rapid, on-site fabric hue inspection on production lines.