Image segmentation method for coal particle size distribution analysis
Graphical abstract
Introduction
In the coal preparation process, particle size distribution (PSD) is very important for determining the load of the classification equipment and the separation effects of particles with extensive classes (Liu et al., 2014). In China, manual screening is currently the most commonly employed technique for PSD analysis in coal preparation plants and laboratories. Manual screening has many disadvantages, such as high labour costs, slow speed, a considerable time lag, and screening is limited by the particle size; all of which prevent the production process from being guided by the results in a timely manner. Accordingly, an online PSD analysis is urgently required for the development of an intelligent coal preparation process.
Image analysis is a very effective method for realising online PSD analysis. For rock and coal particle images with high powder contents, however, the presence of fines may lead to unstable or spurious segmentation results if image segmentation methods are adopted; hence, segmentation methods are not well-suited for online PSD analysis. Many researchers have extracted image texture and other features as machine learning input variables to estimate the PSD (Hamzeloo et al., 2014, Jemwa and Aldrich, 2012, Ko and Shang, 2011, Yaghoobi et al., 2019). The PSD results based on a machine learning method, however, can only be one of several training labels. Its use, therefore, leads to significant error.
When the fine content is low, the coverage of fines on the surfaces of rocks or coal particles is not extensive, and the particle outline is clear. Each particle can therefore be distinctly separated by image segmentation, and the PSD can be obtained based on the axial length, area, and other parameters of heterogeneous particles.
Conventional image segmentation methods are not suitable because of the unique characteristics of coal particle images. In view of the narrow colour range of coal particle images and the uneven grey levels of single coal particles resulting from their reflective surface, methods such as global or local threshold segmentation methods (Cervera Gontard et al., 2011, Chen et al., 2006), clustering based on colour or grey levels (Alush and Goldberger, 2016, Choy et al., 2019; Siu Kai Choy et al., 2017, Horváth, 2020 n.d.<-- -->; Shang et al., 2016), and texture segmentation (Parida & Bhoi, 2017) cannot be employed for single particle segmentation. It is also difficult to detect closed particle contour lines using an edge detection method for segmentation (Gupta and Anand, 2017, Liu et al., 2019, Pratondo et al., 2016) because of specular reflection and the detection of some false edges caused by the different gradient sizes of particle edges. The frame segmentation method, which is based on shape, is not suitable for coal because these particles have irregular shapes (Chen and Wang, 2005, Larsen et al., 2006, Meng et al., 2018, Schorsch et al., 2014, Wu et al., 2019). If the watershed algorithm is directly applied to the coal image, then over-segmentation may result (Micheloni et al., 2005). The concave segmentation method, which is based on polygon approximation (Wang, 1998), only functions well for a few particles that coalesce, and it cannot handle numerous coal particles that have adhered together. The fully convolutional networks (Zhang et al., 2018) and holistically nested convolutional network (Chen et al., 2018) are also incapable of obtaining better segmentation results.
In summary, a single image segmentation method cannot achieve good results for coal particle images; however, the combination of various methods may achieve better results. Zhang et al. (Zhang et al., 2013) segmented the images of coal piles on a belt based on the combination of Hessian matrix edge intensity and watershed segmentation. Although the segmentation results are good, gaps are not split in the segmentation process, thus resulting in segmented particles that are larger than the actual ones.
In this paper, an effective segmentation method that combines the watershed algorithm using gradients with a k-nearest neighbour (KNN) region merging algorithm is proposed for coal particle segmentation. The region merging method effectively prevents over-segmentation, and the convex shell method reduces the number of under-segmented particles.
Section snippets
Samples and equipment
Coal samples with particle sizes ranging from 0.25 to 2.5 mm were obtained from the teetered bed separator feed of the Liuwan coal preparation plant, Shanxi province. To reduce the influence of moisture on image segmentation, the coal samples are dried before the images are captured.
The equipment employed for image acquisition consists of a camera, lens, light source, and shading shelf; the setup is shown in Fig. 1. A high-resolution colour charge-coupled device industrial digital camera (Allied
Method
As shown in Fig. 2, the steps involved in coal particle image segmentation are image pre-processing, gradient transformation, watershed segmentation, KNN region merging, and adhered particle segmentation based on the convex shell method.
Segmentation images
Several images were automatically segmented by the proposed method. For comparison, manual segmentation was performed. The manual segmentation was implemented as follows: the boundary of coal particles was distinguished by human eyes, and it was drawn by hand using a red line. The next PSD calculation process was the same as that used for automatic segmentation. Fig. 10 shows four groups of coal particle images after manual and automatic segmentation. The images on the left are manually
Conclusion
In this paper, an image segmentation method for coal PSD analysis is proposed. It uses gradient watershed for pre-segmentation, employs the KNN algorithm for region merging to avoid over-segmentation, and adopts the convex shell method for segmenting conglutinated particles to avoid under-segmentation. Several images are automatically segmented using the proposed and manual methods. It is found that the difference in the PSDs between the two methods is small, indicating that the proposed method
Declaration of interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (27)
- et al.
A simple algorithm for measuring particle size distributions on an uneven background from TEM images
Ultramicroscopy
(2011) - et al.
Automated cement fragment image segmentation and distribution estimation via a holistically-nested convolutional network and morphological analysis
Powder Technology
(2018) - et al.
Fuzzy bit-plane-dependence image segmentation
Signal Processing
(2019) - et al.
Fuzzy model-based clustering and its application in image segmentation
Pattern Recognition
(2017) - et al.
A hybrid edge-based segmentation approach for ultrasound medical images
Biomedical Signal Processing and Control
(2017) - et al.
Estimation of particle size distribution on an industrial conveyor belt using image analysis and neural networks
Powder Technology
(2014) - et al.
Estimating size fraction categories of coal particles on conveyor belts using image texture modeling methods
Expert Systems with Applications
(2012) - et al.
A neural network-based soft sensor for particle size distribution using image analysis
Powder Technology
(2011) - et al.
An algorithm for analyzing noisy, in situ images of high-aspect-ratio crystals to monitor particle size distribution
Chemical Engineering Science
(2006) - et al.
A weighted edge-based level set method based on multi-local statistical information for noisy image segmentation
Journal of Visual Communication and Image Representation
(2019)