Robust image segmentation technique for rock fragmentation analysis
The file is a zipped PDF document.IntroductionThis paper describes a novel software application that has the capability to autonomously capture images of blasted rock material and analyze them to generate the particle size distribution. The system can also process a batch of rock images captured during a fixed duration of time and produce the overall particle size distribution. The new method has different layers of segmentation modules, which allows the system to respond to a wide range of rock textures and lighting conditions. A new grey-level slicing technique is developed that can perform under a range of illuminating conditions. The Canny-based edge detection technique is used to segment rocks appearing in dark regions.Method OverviewThe main features in the analysis routines of this new software are shown in the figure. The image segmentation part has two main routines: progressive grey-level slicing and multi-level edge detection. The segmented images produced from each routine are combined to produce the final segmented image for analysis. In each process, blob reconstruction and blob reasoning modules are included. The reconstruction routine searches for narrow passages in segmented blobs and reconstructs edges to obtain complete segmentation. The segmented blobs are then analyzed against predefined features to extract properly segmented rocks.Grey-level slicing or binarizing is the most fundamental way of segmenting objects from its background. Binarization segments objects (or pixels) having grey intensities above a given threshold value. When an external lighting device illuminates a pile of rocks, rocks that are closer to the light source and rocks having larger dimensions tend to produce brighter intensities in the image. Such rocks also cast strong shadows around them. The rocks in an image usually appear with varying brightness, and the shadows have varying grey levels. To address this issue, grey-level slicing is performed progressively over a range of intensity values. The binarization function is quite helpful in isolating rocks appearing bright in the image and avoids over segmentation. However, when images are captured under diffused or natural lighting, the rocks may appear with uniform brightness. Under those circumstances, binarization will produce unacceptable segmentation. To avoid this problem, a secondary blob segmentation technique based on edge detection is incorporated.Edge detection algorithms are usually highly sensitive to noise. The effect of noise could be avoided by selecting a higher edge threshold value. However, higher edge thresholding tends to eliminate poorly visible edges in the edge image and would result in incomplete segmentation. On the other hand, a lower edge threshold will cause amplification of noise and result in over-segmentation. To avoid this problem, edge-based segmentation is performed at multiple levels.The proposed image segmentation algorithm has many parameters that can be adjusted to reconfigure the processing system for a variety of applications. In order to provide basic constants, an ‘analysis constant’ file is created. This file will provide the constants such as minimum and maximum particle diameters in centimetres, scale factor (distance to pixel ratio), enabling/disabling fine estimation function, and the region of interest for segmentation. The second important file is given by ‘tuning variables’ where the parameters allow the software system to be reconfigured for a variety of different applications. As an example, the feature limits required for a blob reasoning module can be readjusted depending on rock shapes. To show the validity of the proposed method, three different types of images have been analyzed and their results are shown. The tuning variables are adjusted to obtain optimum performance.