Rock Recognition Using Feature Classification

| A knowledge based approach to nding xed size rocks within an image is described. It is intended to be a building block within a mul-tiresolution system. Each point is hypothe-sised to be a rock and a region surrounding it is labelled using knowledge of rock characteristics. Twelve features are then measured and used to test the hypothesis by means of a combination of thresholding and k-nearest-neighbour classiication. Measurement of rock size distribution or fragmentation has application in the mining industry where sophisticated control systems are used to monitor and control autogenous mills. The use of image processing as a non-invasive measurement technique ooers advantages over conventional techniques such as: undisturbed feed monitoring and repeatble results. Hunter et al 1] gives an overview of image processing techniques for measuring fragmentation. See McDermott and Miles 2], Ivanov et al 3] and Berger 4] for further examples. This paper describes a knowledge based approach to determining the position of xed size rocks within an image. This solution is being used as a building block within a larger multiresolution system for measuring fragmentation using image processing. To take advantage of the multiresolution approach, all processing has been matched to a particular size of object, later referred to as the optimal size. The radius of an optimally sized disc will be referred to as the optimal radius. For a similar approach using neural networks see Crida 5]. The system described here will attemp to detect rocks at each level in a multiresolution image pyramid and compile a list of detected rocks from which the size distribution can be calculated. Although time consuming, this approach to the problem is more robust and accurate than more direct techniques which attempt to measure a size distribution directly by means of chord measurements 6]. In addition the performance is more easily veriied since each rock identiied by the system can be selected and highlighted on an image. The process of detecting optimally sized rock will be referred to as the system. The system described here uses knowledge of rock images to postulate rules which are appropriate when an optimally sized rock is present in the region of interest. In the rst part of the system, it is initially hypothesised that each point in the image is the centre of an optimally sized rock. The rules are then used to determine the extent of a blob in a region …