Rock Grindability Estimation Based On The Quaternion Color Extraction Model

In mineral processing plants it is important to estimate rock composition, size and grindability to improve control of the grinding process. It is well known that variation in ore grindability and size distribution directly affects mills power consumption and throughput. This paper extends a general machine vision approach for on-line estimation of rock mixture composition using color information. Our proposed rock lithological recognition method requires the following steps: division of each image into sub-images, color feature extraction using the Binary Quaternion-Moment-Preserving thresholding technique (BQMP) and support vector machines (SVM) for classification. The BQMP thresholding technique splits the image in two and chooses representatives of each half using the histogram as features. The statistical parameters of the color data can be expressed using the quaternion moments into the representatives. This method can be recursively applied to the sub-images, obtaining 5x2n features in n iterations. Once the feature vector has been computed, each vector is assigned to one of three classes using a classifier. The method was tested on a database containing 20480 sub-images (64x43 pixels) of five ore types as follows: massive sulphide (MS), disseminated sulphide (DS), net textured (NT), gabbro (G), and peridotite (P). These ore types were assigned to three grindability classes: soft (MS), medium (DS and NT) and hard (G and P). The database was divided in 2 subsets: 15360 sub-images used for training and cross validation and 5120 sub-images used for test. The classification accuracy was compared with a method previously published based on a mixture of principal components analysis (PCA) for color, and wavelet texture analysis (WTA) for texture feature extraction. WTA-PCA approach reached a 76% accuracy in test while our BQMP method reached 90% of classification accuracy using only color features. Experimental results show that our proposed method yields excellent results compared to previously published results. Keywords: rock classification, lithological classification, grindability estimation