Prediction Of Concentrate Grade Based On Foam Image Processing

The flotation method is an important method in the beneficiation operation.In the flotation process,the workers guide the production according to the state of the flotation foam combined with their own work experience,and the workload is heavy and subjective.With the development of image processing technology and automation industry,computer vision technology is widely used in industrial automation control production.In this paper,by extracting the image characteristics of gold antimony ore flotation foam,and establishing a concentrate grade prediction model,real-time detection of concentrate grade is realized.The main work of this paper is as follows:(1)Set up an image acquisition device according to the flotation site environment to realize real-time monitoring of flotation foam.According to the characteristics of the flotation foam image,13 types of features are extracted.First,for the rich color features of the flotation foam image,some color component features are extracted in the RGB color space and the HSL color space.Then,to solve the problem that the bubbles in the flotation foam image are difficult to segment,the mark watershed is used for segmentation.The algorithm marks the local maximum value of the flotation foam image and suppresses the excessive segmentation of the bubbles.And extracted bubble size variance,bubble size and other characteristics from the segmented bubble contour.Finally,the SIFT feature matching method is used when extracting the foam velocity feature,and the velocity feature is extracted by calculating the displacement of the feature point.(2)A single feature cannot fully reflect the foam image information,and any combination of multiple features is likely to cause feature redundancy.In this paper,by extracting 13 types of foam image features,the SVM-RFEBPSO concentrate grade prediction model based on feature selection is established.The model first uses the SVM-RFE algorithm to remove some redundant features,reduces the search space of the particle swarm algorithm,and finds the optimal feature subset through the particle swarm algorithm,and finally realizes the concentrate grade prediction through the support vector machine.Experimental results show that through the combination of SVM-RFE algorithm and particle swarm algorithm,the optimal combination of foam image features is achieved,redundant features are reduced,and prediction accuracy is improved.(3)The flotation foam image acquisition process is susceptible to noise interference,which increases the difficulty of feature extraction,and the extraction of multiple features increases the error of feature extraction.This paper also uses the convolutional neural network to extract the characteristics of the foam image and establishes a CNN-SVM concentrate grade prediction model.The experimental results show that in the small sample data of flotation foam images,the performance of the Res Net101-based concentrate grade prediction model is better than the prediction model built by the Alex Net and VGG16 networks.The addition of SVM to the Res Net101 network makes the generalization performance of the model better and the accuracy higher.