Elsevier

Minerals Engineering

Volume 172, 1 October 2021, 107020
Minerals Engineering

Ore image classification based on small deep learning model: Evaluation and optimization of model depth, model structure and data size

https://doi.org/10.1016/j.mineng.2021.107020Get rights and content

Highlights

  • Developing more suitable small deep learning model for ore image classification.

  • Evaluating the gas-coal image classification performance of small deep learning models with different depths.

  • Optimizing the convergence speed and classification accuracy of small deep learning classification models by adding BN layer.

  • Evaluating the gas-coal image classification performance of small deep learning models under different dataset sizes.

Abstract

The ore image classification technology based on deep learning is an effective way to improve the image sensor-based ore sorting classification capability. However, in practice, the image sensor-based ore sorting technique often has the problem of insufficient data, and has not systematically considered the impact of model structure and dataset size on the modeling efficiency and classification performance of deep learning. Therefore, this paper attempts to explore a more suitable small deep learning model for ore image classification by considering the model depth, model structure, and dataset size. Six Convolutional Neural Networks (CNNs) models are established with different depths based on Alex Net and VGG Net and the model structure is optimized by adding BN layer. Taking the gas-coal image dataset as case study, we systematically explore the influence of model depth, model structure, dataset size on the training process efficiency and classification accuracy. Meanwhile, the operational process of coal image classifiers is analyzed visually through the ways of Channel Visualization maps, Heatmaps, Grad-CAM map, and Guided Backpropagation maps.

Introduction

Mineral resources as non-renewable resources, the reserves gradually decrease with mining. At present, the decline of mineral reserves and the increase of gangue treatment capacity put forward new requirements for mineral mining and separation. In order to reduce the high energy consumption of crushing and grinding of block ore, the primary solution is to quickly realize the gangue discharge or pre-separation of the feed materials in the underground or mineral preparation plant.

To solve the above problems, high-tech sensors have been applied to mineral identification and separation tasks instead of manually picking up gangue or wet gangue discharge. At this stage, intelligent dry separation equipment has been applied in underground gangue discharge or ore pre-separation, significantly reducing the energy consumption in the crushing process and reagent consumption in the flotation process, further improving production efficiency and reducing pollution treatment cost. In the current market, the mainstream intelligent ore sorting equipment used into production are based on ray-based sensors, such as γ-ray (Watt and Steffner, 1985), Raman spectrum analysis (Ishikawa and Gulick, 2013), XRT (Robben et al., 2020), XRF (Li et al., 2019). Radiation-based sensors have high classification accuracy and are mostly used to identify and separate large ore blocks. However, it has the disadvantages of high cost and high radiation, limiting the application and development of intelligent ore detection technology to some extent.

One of the potential alternatives to solve these limitations are machine vision-based ore sorting, which has the advantages of low cost, high efficiency, no radiation, easy installation. Mineral apparent properties are mainly obtained through optical components to identify mineral attributes (Zhang et al., 2020a). Presently, most of the studies focus on machine learning classification algorithms, which are mainly divided into supervised learning algorithms and unsupervised learning algorithms. Among them, the supervised learning algorithm has the most comprehensive application range. It preprocesses the input mineral images and builds classification models to complete mineral separation. However, it is unavoidable that there is a bottleneck in developing and applying machine learning technology, since the high-efficiency application of machine learning-based ore image classification needs to be based on high-resolution images. Firstly, high-resolution images make image acquisition more difficult, especially in a few harsh environments. Additionally, high-resolution images mean longer recognition time, which will reduce the ore sorting efficiency.

In recent years, deep learning technology has made outstanding achievements in many fields. It uses the Convolutional Neural Networks (CNNs) calculation to reduce the dependence on image resolution, simplify the training process, and improve the classification speed, which urges mineral sorting researchers to use it in ore sorting tasks. The results indicate that the deep learning algorithm has higher classification efficiency and accuracy than the machine learning algorithm for mineral image classification, which proves its potential to be applied in industrial practice as the core of ore sorting technology. However, with the deepening of the network and the increasing complexity of the network, dozens of layers of deep learning network make model building and model training more difficult, which are reflected in the following three aspects: (1) high hardware requirements; (2) long model training time; (3) the quantity of image dataset is not enough to support the needs of model training. One possible solution is transfer learning, which combines pre-training models that extract basic features with actual industrial tasks, effectively reducing over-fitting problems, and solving the problems of high hardware requirements, long training times, and insufficient data (Pan and Yang, 2010). Although transfer learning technology solves the above limitations to a certain extent, the non-targeted basic features limit the rising space of the classification accuracy.

Hence, we systematically evaluate and optimize the training efficiency and classification accuracy of ore image by considering the model depth, model structure, dataset size of six small deep learning models established under the guidance of Alex Net and VGG Net. This paper is of great significance for the selection of small deep learning classification models for ore images with different dataset size.

Section snippets

Related works

Imaging technology and computer engineering have witnessed tremendous progress, promoting the development of ore sorting technology based on visible light. The information obtained by optical components is mainly digital signal images. Therefore, efficient recognition and classification of digital signal images are the centers of the application of visible light-based machine vision technology to mineral image classification.

Image segmentation

The color threshold segmentation process consists of two types of segmentation steps: internal particle segmentation (Fig. 1(a)) and edge particle segmentation (Fig. 1(b)), as shown in Fig. 1.

In the color threshold segmentation process, the original image is first segmented by the target area and the background area (Eq. (1)).fx,y,z=0,0,0y-x>Mx<Mx,y,zotherwhere f (x, y, z) is the pixel value of any point in the RGB color image, x is the pixel value of the R component, y is the pixel value of

Material preparation

This study taken the gas coal in China as the experimental object. In order to minimize the influence of errors in the classification process and effectively obtain the classification results of gas coal images, 13–25 mm particle size was selected as the experimental object by artificial screening and was divided into two categories according to the density classes: <1.6 g/cm3 and >1.6 g/cm3. Firstly, this classification method can be applied to the direct sales of granular coal products in

Conclusion & outlook

In order to further promote the application potential of image classification technology based on deep learning in ore sorting and improve the classification accuracy of mineral images, this paper makes an in-depth exploration of a more suitable small deep learning model for ore image classification by considering the model depth, model structure, and dataset size. The specific conclusions are as follows:

  • (1)

    In terms of the effect of model depth on the classification accuracy of mineral images, the

CRediT authorship contribution statement

Yang Liu: Data curation, Writing - original draft. Zelin Zhang: Conceptualization, Methodology, Software. Xiang Liu: Software, Validation. Lei Wang: Writing - review & editing. Xuhui Xia: Supervision.

Declaration of Competing Interest

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.

Acknowledgement

The authors would like to thank the supported by the National Natural Science Foundation of China (No. 51604196 and 51805385), and the Key Research and Development project of Hubei Province (No. 2020BAA024 and 2020BAB047).

References (50)

  • H. Izadi et al.

    An intelligent system for mineral identification in thin sections based on a cascade approach

    Comput. Geosci.

    (2017)
  • B.T. Le et al.

    Coal analysis based on visible-infrared spectroscopy and a deep neural network

    Infrared Phys. Technol.

    (2018)
  • J. Li et al.

    Real-time Detection of Steel Strip Surface Defects Based on Improved YOLO Detection Network

    IFAC-PapersOnLine

    (2018)
  • M. Massinaei et al.

    Machine vision based monitoring and analysis of a coal column flotation circuit

    Powder Technol.

    (2019)
  • C.E. Metz

    Basic principles of ROC analysis

    Semin. Nucl. Med.

    (1978)
  • A. Mollajan et al.

    Improving pore type identification from thin section images using an integrated fuzzy fusion of multiple classifiers

    J. Nat. Gas Sci. Eng.

    (2016)
  • G. Montes-Atenas et al.

    Predicting bubble size and bubble rate data in water and in froth flotation-like slurry from computational fluid dynamics (CFD) by applying deep neural networks (DNN)

    Int. Commun. Heat Mass Transf.

    (2016)
  • A. Nasiri et al.

    Image-based deep learning automated sorting of date fruit

    Postharvest Biol. Technol.

    (2019)
  • C. Robben et al.

    X-ray-transmission based ore sorting at the San Rafael tin mine

    Miner. Eng.

    (2020)
  • M. Sadeghiamirshahidi et al.

    Application of artificial neural networks to predict pyrite oxidation in a coal washing refuse pile

    Fuel

    (2013)
  • N. Sharma et al.

    An Analysis of Convolutional Neural Networks for Image Classification

    Procedia Comput. Sci.

    (2018)
  • X. Wang et al.

    Process working condition recognition based on the fusion of morphological and pixel set features of froth for froth flotation

    Miner. Eng.

    (2018)
  • J.S. Watt et al.

    Dual energy gamma-ray transmission techniques applied to on-line analysis in the coal and mineral industries

    Int. J. Appl. Radiat. Isot.

    (1985)
  • J. Woo et al.

    Dynamic model identification of unmanned surface vehicles using deep learning network

    Appl. Ocean Res.

    (2018)
  • Y. Xiong et al.

    Mapping mineral prospectivity through big data analytics and a deep learning algorithm

    Ore Geol. Rev.

    (2018)
  • Cited by (75)

    View all citing articles on Scopus
    1

    Contributes equally to this work.

    View full text