Scholarlink

<p>Effective monitoring of industrial flotation circuits is critical in mineral processing, as it directly affects separation quality and operational costs. Current manual methods, which rely on visual inspection, are inconsistent and subjective, leading to unreliable process control. While recent automated approaches employ image classification to recognize flotation working conditions, they neglect spatio-temporal relationships between video frames. This study proposes a video-based classification framework for flotation froth image sequences that integrates deep learning networks (InceptionV3 and VGG16) to extract spatial features. These features are processed through a Principal Component Analysis module to retain significant information from each frame while discarding redundant data. The refined features are then fed into a Gated Recurrent Unit network, which analyzes temporal dependencies. The network’s hidden state is fed to a classification head to classify the froth image. The study evaluates a four-class flotation froth video dataset comprising 2,386 videos, each with 12 frames, recorded at an industrial flotation plant. The four classes cover stable froth, slightly denser froth, broken froth with impurities, and sparse, uneven froth. Results achieve 0.982 ± 0.002 accuracy and strong scores across all metrics, surpassing the performance of existing techniques.</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

Enhanced Multi-Branch Spatio-Temporal Fusion with PCA-GRU for Flotation Froth Image Sequence Classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月16日

查看全文
  • Khalid A. Abouda,
  • Degang Xu,
  • Wail M. Idress,
  • Nahlah Algethami,
  • Noura Albarakati,
  • Laeeq Aslam

摘要

Effective monitoring of industrial flotation circuits is critical in mineral processing, as it directly affects separation quality and operational costs. Current manual methods, which rely on visual inspection, are inconsistent and subjective, leading to unreliable process control. While recent automated approaches employ image classification to recognize flotation working conditions, they neglect spatio-temporal relationships between video frames. This study proposes a video-based classification framework for flotation froth image sequences that integrates deep learning networks (InceptionV3 and VGG16) to extract spatial features. These features are processed through a Principal Component Analysis module to retain significant information from each frame while discarding redundant data. The refined features are then fed into a Gated Recurrent Unit network, which analyzes temporal dependencies. The network’s hidden state is fed to a classification head to classify the froth image. The study evaluates a four-class flotation froth video dataset comprising 2,386 videos, each with 12 frames, recorded at an industrial flotation plant. The four classes cover stable froth, slightly denser froth, broken froth with impurities, and sparse, uneven froth. Results achieve 0.982 ± 0.002 accuracy and strong scores across all metrics, surpassing the performance of existing techniques.