Scholarlink

<p>The pharmaceutical manufacturing industry faces critical challenges in maintaining consistent quality control during high-speed tablet production, where existing tracking methods suffer from identity switches, temporal inconsistencies, and inaccurate counting mechanisms. This paper presents the Multi-Scale Pharmaceutical Defect Tracker with Adaptive Confidence and Counting (MS-PDTAC), a comprehensive real-time system for defect detection, tracking, and counting in pharmaceutical tablet production lines. The system employs a dual-stream concurrent processing architecture integrating YOLOv8-based object detection with sophisticated tracking mechanisms and self-supervised learning capabilities. The primary processing pipeline implements multi-stage false detection filtering through geometric validation, non-maximum suppression, and temporal consistency checking, followed by multi-scale histogram feature extraction at four resolution levels to capture defect characteristics across varying scales. Detection-to-track association employs Hungarian algorithm optimization with composite cost matrices that integrate spatial distance, appearance similarity, and confidence quality components, enhanced by a spatio-temporal attention mechanism that dynamically weights these factors based on trajectory prediction, temporal consistency, and feature matching reliability. State estimation utilizes Kalman filtering for position prediction and exponential smoothing for feature refinement, while defect state lifecycle management implements tentative-to-active state transitions, ensuring only validated defects contribute to final counts. Multi-method line crossing detection combines direct position monitoring, trajectory history analysis, and area-based detection to achieve robust counting accuracy. A concurrent self-supervised learning stream continuously adapts confidence thresholds and feature weights through statistical feedback mechanisms operating every 50 frames, enabling real-time system optimization without manual intervention. Comprehensive evaluation on custom pharmaceutical datasets demonstrates perfect counting accuracy (100%) with zero absolute count error across all seven defect categories (black dots, broken tablets, color mismatches, cracks, dual caps, empty capsules, and foreign particles), significantly outperforming the base methodology that suffers from over-counting, false positives, and identity switching issues.</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

Multi-Scale Trajectory Tracking of Pharmaceutical Defects with Adaptive Learning and Precision Counting

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Ajantha Vijayakumar,
  • Joseph Abraham Sundar Koilraj,
  • Muthaiah Rajappa

摘要

The pharmaceutical manufacturing industry faces critical challenges in maintaining consistent quality control during high-speed tablet production, where existing tracking methods suffer from identity switches, temporal inconsistencies, and inaccurate counting mechanisms. This paper presents the Multi-Scale Pharmaceutical Defect Tracker with Adaptive Confidence and Counting (MS-PDTAC), a comprehensive real-time system for defect detection, tracking, and counting in pharmaceutical tablet production lines. The system employs a dual-stream concurrent processing architecture integrating YOLOv8-based object detection with sophisticated tracking mechanisms and self-supervised learning capabilities. The primary processing pipeline implements multi-stage false detection filtering through geometric validation, non-maximum suppression, and temporal consistency checking, followed by multi-scale histogram feature extraction at four resolution levels to capture defect characteristics across varying scales. Detection-to-track association employs Hungarian algorithm optimization with composite cost matrices that integrate spatial distance, appearance similarity, and confidence quality components, enhanced by a spatio-temporal attention mechanism that dynamically weights these factors based on trajectory prediction, temporal consistency, and feature matching reliability. State estimation utilizes Kalman filtering for position prediction and exponential smoothing for feature refinement, while defect state lifecycle management implements tentative-to-active state transitions, ensuring only validated defects contribute to final counts. Multi-method line crossing detection combines direct position monitoring, trajectory history analysis, and area-based detection to achieve robust counting accuracy. A concurrent self-supervised learning stream continuously adapts confidence thresholds and feature weights through statistical feedback mechanisms operating every 50 frames, enabling real-time system optimization without manual intervention. Comprehensive evaluation on custom pharmaceutical datasets demonstrates perfect counting accuracy (100%) with zero absolute count error across all seven defect categories (black dots, broken tablets, color mismatches, cracks, dual caps, empty capsules, and foreign particles), significantly outperforming the base methodology that suffers from over-counting, false positives, and identity switching issues.