【作者】 张蕾；

【导师】 陈宁； 彭海良；

【作者基本信息】 中南大学 ， 电子与通信工程， 2014， 硕士

【摘要】 近年来,数字图像处理技术在矿物浮选过程中得到了广泛的研究与应用。通常,研究浮选泡沫表层特征参数的变化,能够对泡沫浮选工况进行一定的识别。但是,由于浮选工艺现场环境恶劣,导致浮选泡沫图像品质低,从而影响对浮选泡沫图像特征提取的准确性,最终影响浮选泡沫图分类识别的精确率。为了降低工业现场对图像品质的影响,提高浮选泡沫图分类准确率,本文提出了基于向量空间模型(Vector Space Model,VSM)的泡沫图像纹理特征提取和分类方法。(1)本文针对泡沫图像特征提取受图像品质的影响,提出了基于VSM的纹理特征提取方法。该方法通过对数据库图像进行合理分块,提取各分块的颜色共生矩阵(Color Co-occurrence Matrix, CCM)纹理特征,对所有的CCM特征向量用模糊C均值聚类得到图像的CCM纹理数据表,并用设计的相对TF-MI(Term frequency-Mutual information)权重因子对CCM纹理数据表进行加权,再将对单幅图像各个分块的CCM特征向量与加权后的CCM纹理数据表进行对照并标识每个分块的状态,统计各个状态出现的频次,最后得到图像CCM纹理特征的向量表示称为CCM纹理特征的词袋表示。(2)分别采用BP (Back Propagation)和LVQ (Learning Vector Quantization)神经网络模型对获得的词袋数据分类处理。针对BP网络稳定性受隐含层节点影响、LVQ网络精度不足训练时间过长等问题,设计了BP-LVQ网络组合分类模型。该模型为采用三层网络结构,底层选取4组每组2个初始值不同的BP网络,中层为4个结构相同的LVQ,顶层为1个LVQ网络。其中底层各组网络采用不同的隐含层节点数,将每组的结果取均值输入到对应的中层LVQ网络,中层的4个LVQ网络再将输出结果按权线性组合输入顶层的LVQ网络,由顶层LVQ网络输出最终分类结果。(3)将所提的基于VSM的特征提取和分类方法应用在铝土矿浮选过程中。结果表明：本文所提出的浮选泡沫特征提取和分类方法能提高分类准确性,能够对浮选过程智能控制提供参考。更多还原

【Abstract】 In recent years, digital image technology has been widely researched and applied in the process of mineral flotation. Usually, the flotation condition can be distinguished by studying the variation of the foam surface characteristic parameters. However, the environment of the flotation scene is always very bad, result in the quality of the froth image is low and affecting the accuracy of flotation bubble image feature extraction. It will ultimately affect the precision of the flotation foam figure classification recognition rate. In order to reduce the bad effects of industrial field on image quality, improve the accuracy of froth image classification, this paper proposed a method of froth image texture feature extraction and classification based on vector space model (VSM).(1) In this paper, the research status of image texture extraction and classification methods are summarized at first. In view of the bubble image feature extraction is affected by image quality, method of texture feature extraction based on VSM is proposed. This method blocks image of database reasonably, and extracts the color co-occurrence matrix (CCM) characteristics of each block. The image’s CCM texture data table is got by using the fuzzy c-means clustering on all CCM feature vectors and weighted by using the relative TF-MI (Term frequency-Mutual information) weighting factor. Then, the CCM texture vector of each block is contrasted with the weighted data table and state of each block is marked. At last, the frequency of each state is censused and the image is represented by CCM texture vector called CCM texture feature word bag representation.(2) In order to use the word bag of CCM texture to reflect the picture’s feature type, this paper applied BP and LVQ neural net model to classify the word bag data. The CCM texture word bag vector set and the set of image are used to study the problem of image classification. Because of BP network stability is affected by the hidden layer nodes, LVQ network’s precision is low and raining time is too long, this paper designed a classification model named BP-LVQ reliable net combination model. This model used three-layer network, chose8BP networks as bottom layer to overcome the low precision of LVQ network. The8BP networks are divided into4groups and each group used different number of hidden layer nodes. Then, the mean value of each group’s result is used as input to the middle layer and the middle lay includes4same structure LVQ networks. It can synthesize the output results of different BP networks in bottom layer and the affection of hidden layer nodes can be solved by this model too. At last, the4outputs of LVQ network in top layer according to reliable linear combination algorithm and the top LVQ network output the final classification result.(3)Use the method texture feature extraction and classification based on the VSM application in aluminum froth floatation. The result show that the flotation froth feature extraction and classification method can improve the classification accuracy, and can provide reference for intelligent control of flotation process.更多还原