Elsevier

Particuology

Volume 56, June 2021, Pages 163-170
Particuology

Image segmentation method for coal particle size distribution analysis

https://doi.org/10.1016/j.partic.2020.10.002Get rights and content

Highlights

  • Image segmentation is utilised for coal particle size distribution analysis.

  • A watershed algorithm with gradient is employed for preliminary segmentation.

  • The k-nearest neighbour algorithm is utilised to merge small pieces to particles.

  • The convex shell method is introduced to segment adhered particles.

Abstract

Particle size distribution is extremely important in the coal preparation industry. It is traditionally analysed by a manual screening method, which is relatively time-consuming and cannot immediately guide production. In this paper, an image segmentation method for images of coal particles is proposed. It employs the watershed algorithm, k-nearest neighbour algorithm, and convex shell method to achieve preliminary segmentation, merge small pieces with large pieces, and split adhered particles, respectively. Comparing the automated segmentation using this method with manual segmentation, it is found that the results are comparable. The size distributions obtained by the automated and manual segmentation methods are nearly identical, and the standard deviation is less than 3%, indicating good reliability. This automated image segmentation method provides a new approach for rapidly analysing the size distribution of coal particles with size fractions defined according to consumer requirements.

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)

Cited by (0)

View full text