Scholarlink

<p>The identification and parameter extraction of ore and rock particles pose challenges due to the complex characteristics of particles, aggregation issues, and poor image quality. Common instance segmentation methods often struggle with over-segmentation or under-segmentation at particle edges. To address these challenges, four leading instance segmentation networks were developed and evaluated based on convergence, accuracy, robustness, and generalization. The results of parameter extraction from each network were utilized to validate their performance. The research findings suggest that: (1) the you only look at coefficients (YOLACT) network has low completeness and segmentation effectiveness, while the segmenting objects by locations (SOLO) network tend to over-segment at particle edges. Conversely, the mask region-based convolutional neural network (Mask R-CNN) network exhibits tendencies of both over-segmentation and under-segmentation at particle edges. In contrast, the masked-attention mask transformer (Mask2Former) network achieves detection and segmentation accuracies of 0.878 and 0.904, respectively, without notable issues of under-segmentation or over-segmentation. (2) The Mask2Former model demonstrates superior performance compared to other networks concerning convergence, accuracy, robustness, and generalization. Specifically, the <i>Box</i>_<i>mAP</i> metric exceeds that of other models by a margin ranging from 0.017 to 0.431, while the <i>Segm</i>_<i>mAP</i> metric shows an improvement of 0.039–0.408. (3) The Mask2Former model displays better consistency with actual data in detecting rock particles number and gradation statistics, with maximum error and average error in particle size proportions being 2.37% and 1.38%, respectively.</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

Intelligent Identification and Gradation Calculation of Aggregated Ore and Rock Particles Based on Instance Segmentation

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Zongsheng Dai,
  • Hao Sun,
  • Lishan Zhao,
  • Xuan Qin,
  • Lichang Wei,
  • Xinmin Liang,
  • Tingting Chen,
  • Junze Jia

摘要

The identification and parameter extraction of ore and rock particles pose challenges due to the complex characteristics of particles, aggregation issues, and poor image quality. Common instance segmentation methods often struggle with over-segmentation or under-segmentation at particle edges. To address these challenges, four leading instance segmentation networks were developed and evaluated based on convergence, accuracy, robustness, and generalization. The results of parameter extraction from each network were utilized to validate their performance. The research findings suggest that: (1) the you only look at coefficients (YOLACT) network has low completeness and segmentation effectiveness, while the segmenting objects by locations (SOLO) network tend to over-segment at particle edges. Conversely, the mask region-based convolutional neural network (Mask R-CNN) network exhibits tendencies of both over-segmentation and under-segmentation at particle edges. In contrast, the masked-attention mask transformer (Mask2Former) network achieves detection and segmentation accuracies of 0.878 and 0.904, respectively, without notable issues of under-segmentation or over-segmentation. (2) The Mask2Former model demonstrates superior performance compared to other networks concerning convergence, accuracy, robustness, and generalization. Specifically, the Box_mAP metric exceeds that of other models by a margin ranging from 0.017 to 0.431, while the Segm_mAP metric shows an improvement of 0.039–0.408. (3) The Mask2Former model displays better consistency with actual data in detecting rock particles number and gradation statistics, with maximum error and average error in particle size proportions being 2.37% and 1.38%, respectively.