基于复原图像特征与深度视觉特征融合的锑粗选异常工况识别
Fault condition recognition based on restored image features and deep visual features
摘要点击 48  全文点击 33  投稿时间:2019-05-29  修订日期:2019-10-14
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2019.90392
  2020,37(6):1207-1217
中文关键词  泡沫浮选  工况识别  模糊核零范数  灰度平均梯度差  迁移学习  InceptionResNetV1  XGBoost
英文关键词  froth flotation  fault condition recognition  blur kernel zero norm  gray mean gradient deviation  transfer learning  InceptionResNetV1  XGBoost
基金项目  国家杰出青年科学基金项目(61725306), 国家自然科学基金广东联合基金重点项目(U1701261)资助.
作者单位E-mail
保江 中南大学 bj18229949143@csu.edu.cn 
谢永芳 中南大学  
刘金平 湖南师范大学 ljp202518@163.com 
唐朝晖 中南大学  
艾明曦 中南大学  
中文摘要
      本文针对锑粗选异常工况下泡沫层高度改变, 导致位置固定的工业相机采集到的泡沫图像存在离焦模糊 的问题, 提出了一种基于泡沫复原图像特征和深度视觉特征融合的锑粗选异常工况识别方法. 该方法首先通过分 析不同工况下模糊泡沫图像的特点, 采用基于L0和L2正则项的模糊核估计方法提取了模糊核零范数特征, 再利用 L–R算法提取了灰度平均梯度差特征. 其次, 为了往更深层次挖掘异常工况下模糊泡沫图像的特征信息, 本文采用 迁移学习方法, 基于InceptionResNetV1深度神经网络, 利用大量泡沫图像数据对深度神经网络进行微调, 进而提取 泡沫图像的深度视觉特征. 最后, 基于XGBoost机器学习算法, 先对高维视觉特征进行降维, 再融合复原图像特征与 深度视觉特征, 对不同工况下的泡沫图像进行分类识别. 锑粗选实验结果表明, 该方法能够有效降低样本错分率, 提 高锑粗选异常工况识别率.
英文摘要
      In the antimony flotation process, an industrial camera installed at the fixed position can not accurately focus the froth images when the froth layer lever changes, which causes the defocus blurred problems of the froth images and even the fault conditions. A fault condition recognition method based on the restored image features and the deep visual features is proposed to deal with the defocus blurred froth images in this paper. Firstly, the characteristics of blurry froth images under different working conditions are analyzed. The blur kernel estimation method based on L0 and L2 regular terms is utilized to extract the blur kernel zero norm feature while the L–R algorithm is used to extract the gray mean gradient deviation feature. Secondly, to mine the deeper feature information of blurry froth images under fault conditions, this paper uses a transfer learning method based on InceptionResNetV1 deep neural network, which uses a large amount of froth image data to fine-tune the deep neural network and extract the deep visual features of the froth image. Finally, based on the XGBoost machine learning algorithm, the high-dimensional visual features are reduced, and then the restored image features and deep visual features are combined to classify and identify the froth images under different working conditions. The results of antimony rougher experiments show that the proposed method can reduce the sample misclassification rate effectively and improve the recognition rate of the fault conditions.