MRMR和SSGMM联合分类模型的煤泥浮选系统药况图像识别

Recognition of reagent dosage condition image for coal flotation system based on joint classification model of MRMR and SSGMM

中文关键词  煤泥浮选泡沫  加药状况  机器视觉  图像特征提取和选择  半监督学习  联合分类模型

英文关键词  coal flotation froth  reagent dosage condition  machine vision  feature extraction and selection  semisupervised learning  joint classification model

中文摘要

      为解决煤泥浮选过程依靠工人肉眼识别泡沫特征来调节药剂用量, 造成药剂浪费, 产品质量不合格的问题, 提出一种MRMR和SSGMM联合分类模型的药况图像识别方法. 针对泡沫图像的形态、纹理、颜色特征与泡沫类别 具有不同程度的相关性. 将精煤灰分作为泡沫的类别信息, 利用最大相关最小冗余(MRMR)算法筛选最优特征; 针 对传统的高斯混合模型(GMM)在聚类时, 存在结果需人为判断实现分类的问题, 通过引入少量已知加药状况下的 泡沫图像特征样本对其改进, 构建半监督高斯混合模型(SSGMM)泡沫图像聚类器. 将优选的且具有少量先验标签 信息的多维泡沫图像特征融合到SSGMM聚类模型中, 利用少量的标记样本引导聚类, 并将其标签信息映射给聚类 结果实现自动分类. 实验表明, 这种联合分类模型提高了泡沫识别的准确性, 为药剂用量的准确控制与精煤产品质 量提供了关键技术支持.

英文摘要

      In order to solve the problem that the coal flotation process depends on the naked eyes of the workers to identify the froth features to adjust the dosage of reagent which results in the waste of reagents and the unqualified product, an recognition method of reagent dosage condition image based on joint classification model of MRMR and SSGMM is proposed. With respect to the different degrees of correlations between the morphology, texture and color features of the froth image and the froth classification, the ash content of clean coal is taken as the classification information, the optimal froth image features are screened out by maximal-relevance-minimal-redundancy (MRMR) algorithm; aiming at the problem that the results need to be judged artificially to realize the classification when the clustering of traditional Gaussian mixture model (GMM), it is improved by introducing a small number of froth image feature samples under the condition of known reagent dosage, a semi-supervised Gaussian mixture model (SSGMM) cluster is constructed. The optimal multi-dimensional froth image features with a small amount of prior label information are integrated into the SSGMM clustering model, the clustering is guided by labeling samples, and their label information is mapped to the clustering result, so that the automatic classification is realized. The experimental results show that the accuracy of froth recognition has been improved by this kind of joint classification model, and the key technical support has been provided for the accurate control of the dosage of reagent and the quality of clean coal products.