1. Introduction

Laser-induced breakdown spectroscopy (LIBS) is a convenient and efficient elemental analysis technology. It has been widely used in many fields, such as metals [1–4], mineral [5], rocks [6,7], and water [8,9] etc. Traditional classification methods can be divided into two categories: full-spectra intensities method (FSIM) and multi-spectral lines intensities method (MSLIM) in LIBS analysis. In-depth research and extensive applications on these methods have been reported in recent years. For example, Gottfried et al. [10]studied the classification of simulant residues based on full-spectra feature model and the intensity/ratio model, and their classification accuracy were 54.8% and 47.4%. Ji et al. [11] adopted LIBS assisted FSIM with partial least squares discriminant analysis (PLS-DA) algorithm for Tegillarca granosa discrimination, the classification accuracy was less than 30%. Zhang et al. [12] employed LIBS coupled the FSIM with support vector machine (SVM) and radial basis function neural network (RBFNN) algorithm to classified the coffee. The classification accuracies of the two methods reached about 83.59 and 82.03%, respectively. To improve the generalization performance of the algorithm, some researchers studied the line selection method of MSLIM. For example, Kong et al. [13] proposed three data reduction methods for LIBS, including background normalization, the internal standard method, and the standard normal variate. N. Labbé et al. [14] proposed multivariate analysis method to extract information from LIBS spectral data. In our previous work [15,16], LIBS assisted MSLIM with decision tree (DT), random forest (RF), PLS-DA, linear discriminant analysis (LDA) and SVM algorithms on rice geographic origins classification, and the classification accuracies were 86.8%, 96.3%, 96.8%, 98.6%, and 99.2%, respectively.

Above researches proved the feasibility of FSIM and MSLIM in LIBS classification, but their classification accuracies for some complex samples with uneven surface such as simulant residues [10] and Tegillarca granosa [11] and different varieties of rice samples still can’t meet some important application requirements. Taking rice sample as an example, the annual output of Wuchang high-quality rice in its origin is about 714,000 tons (official data in 2014), but this variety of rice on the market has reached 10 million tons. The identification and classification of rice producing areas or varieties by LIBS technology can effectively prevent food safety [17,18] and seed adulteration problems [19,20] as described above. The reason for unsatisfactory classification performance should be the feature extraction way of the traditional classification methods cannot effectively extract the features difference between samples [21]. Taking MSLIM and FSIM as examples, their classification features are intensities of multiple spectral lines and full-spectral lines, respectively. Both features are one-dimensional with relatively weak characterization ability. However, the inherent relationships between multi-dimensional spectral intensities can also be used to characterize the samples, but few research have referred on this.

The histogram of oriented gradient (HOG) [22] feature is a kind of feature descriptor for images. It has been widely used in computer vision [23] and image pattern recognition [24] due to its powerful ability on pedestrian detection [25]. Its feature extraction of LIBS images can indirectly realize the extraction of multi-dimensional spectral features.

In this work, a classification method IHFM, which combined with image features and SVM algorithm was proposed. A classification of 24 rice samples was conducted to verify the proposed method. To test the universal applicability of IHFM, classification experiment were also carried out on a variety of plastic, steel, rock and mineral samples.

2. Experimental setup and samples

2.1 Experimental setup

The experimental setup of this study is shown in Fig. 1. A Q-switched Nd: YAG laser (Quantel Brilliant, energy: 40 mJ/pulse, wavelength: 532 nm, frequency: 10Hz, pulse width:8ns) was focused at 1.25 mm below the target surface by a 100 mm focal length lens togenerate plasma. The plasma emission was collected by a collector and then transmitted to a spectrometer (Andor Tech., Mechelle 5000, spectral range from 200 to 950 nm with a resolution of λ/Δλ = 5000, wavelength accuracy of ± 0.05 nm) through an optical fiber. The light signal is converted into electrical signal by an ICCD camera (Andor Tech., iStar DH-334T) and output to computer.

 

Fig. 1 LIBS experimental setup.

Download Full Size | PDF

The samples were placed on a 3-axis displacement stage in this experiment. The scanning area of the laser was the upper surface of the sample, with a diameter of about 40.98mm (the edge thickness of the container is ignored) and an area of 1318.3mm². To collect stable spectra, the displacement stage moved at a speed of 4 mm/s. The time delay between laser pulse signal and spectrometer acquisition was 1.5 μs, and gate width was 3 μs. To reduce spectral fluctuations, 50 pulses were accumulated for each spectra, and 100 spectra were collected for each sample.

2.2 Samples

In experiments, 24 different varieties rice samples were bred by different research institutes in China, as shown in Table 1.

Table 1. Rice samples information list

View Table | View all tables in this article

There is parental relationship among some of the varieties in the sample. For example, both YX725 and YX3728’s male parents are YX 1A, and the female parent is MiHui series varieties. The male parents of JR 372, SR 16, TR 99 and LJ 8 are all Zhen Dao 88. Due to the similarity of the parental relationship of each samples, it is difficult to classify them by the existing classification methods.

3. Images HOG feature extraction and optimization of LIBS parameters

3.1 Extraction procedure and principle

 

Fig. 2 Schematic diagram of characteristics spectrum conversion to characteristics images

Download Full Size | PDF

3.2 Optimization of image column number and image grayscale coefficient

 

Fig. 3 (a). The relationship between the number of image columns and the classification accuracy (b). The relationship between grayscale coefficient and the classification accuracy.

Download Full Size | PDF

4. Results and discussion

4.1 Comparison of IHFM and FSIM for rice samples classification

As can be seen from the section 2.1 and 3.1(1), a total of 2400 spectra and images would be used for the classification. During the classification, 1200 spectra or images among them were used for training the SVM model [28], and the rest 1200 spectra were used for testing. In addition, the SVM algorithm here and its parameters optimization was implemented by the LIBSVM [29–31] tool in the Matlab environment. The classification results of IHFM and FSIM for rice samples were shown in Fig. 4. Classification accuracies of every sample were listed in Table 2.

 

Fig. 4 Comparison of classification results using FSIM (a) and IHFM (b).

Download Full Size | PDF

Table 2. Classification accuracies of FSIM and IHFM for 24 varieties of rice samples

View Table | View all tables in this article

From Fig. 4, the coincidence rate between the predicted label and the actual label using IHFM (Fig. 4(b)) was bigger than that of FSIM (Fig. 4(a)). From Table 2, the classification accuracies of the IHFM were obviously better than those of the FSIM for 24 varieties of rice samples and it increased from 60.25% of FSIM to 81.00% of IHFM.

To further verify the classification accuracies of rice samples, a 10-fold cross-validation was adopted in this work. The results of the cross-validation were 60.25% of FSIM and 76.33% of IHFM, respectively.

The main reason that the classification accuracy of IHFM was better than that of FSIM on rice classification may as follows: when HOG features was extracted from LIBS images in IHFM, it essentially obtained the gradient direction relationship of the multi-dimensional spectral intensities, but only one-dimensional spectral intensities was extracted in FSIM. Due to the weak stability (average RSDs of matrix elements C, H, O were over 31.43%) of the LIBS spectra of these rice samples, and the characterization ability of one-dimensional spectral intensity for each sample were weaker than those of the multi-dimensional spectra, so the classification accuracies of FSIM were relatively poor.

The correlation between the characteristic matrices of the samples can indirectly reflect the feasibility of classification. Generally, the higher the correlation of characteristic matrices, the more difficult the classification, and vice versa [32]. The correlation between the characteristic matrices X and Y was expressed by the correlation coefficient ρ_XY, which was defined as shown in Eq. (2) and Eq. (3).

In the above equations, ${\rho }_{XY}$ and $Cov(X,Y)$are the correlation coefficient and the covariance respectively. ${\sigma }_X$ and ${\sigma }_Y$ are the standard deviations, and ${\mu }_X$ and ${\mu }_Y$ are the means of the characteristic matrices.

Taking the GD1 and ZD2 rice samples as examples, the correlation coefficient ${\rho }_{\text{IHFM}}$ between the image HOG feature matrices and ${\rho }_{FSIM}$ between the full-spectrum characteristic matrices are 0.8527 and 0.9969, respectively. That is to say, ${\rho }_{\text{IHFM}}$ < ${\rho }_{FSIM}$, so the classification accuracy of IHFM is better than FSIM under same algorithm and their respective optimal parameters.

4.2 Universality test for IHFM

To verify the universality of IHFM, further classification tests were carried out on a variety of plastic [33], steel, rock and mineral samples. Categories numbers of the plastic, steel, rock and mineral samples were 20, 18, 15 and 16, respectively. The number of collected spectra of these samples was 150, 100, 100 and 50 respectively. The values of gray parameters and column parameters were fixed to 10 and 183 for plastic, 1 and 64 for steel, 100 and 64 for rock, 100 and 64 for mineral, respectively.

Figure 5 shows that the classification accuracies of IHFM has reached a high level for all the above samples. Among them, the classification accuracies are more than 99% for both plastic and steel samples, and 97.47% and 91.25% for rock and mineral samples.

 

Fig. 5 Classification result comparison of FSIM and IHFM for other samples.

Download Full Size | PDF

To further verify the classification accuracies of these samples, here is the same as the rice classification with a 10-fold cross-validation. For plastics, steel, rocks and minerals, the cross-validation results of FSIM were 99.44, 97.11, 96.67 and 89.00%, respectively, and the results of IHFM were 99.60, 98.78, 97.60 and 91.50%, respectively.

It was shown that IHFM could achieve good classification results not only for the rice samples of different varieties, but for those samples such as plastics, steels, rocks and minerals. The above indicates that, IHFM has good universality.

5. Conclusions

In summary, IHFM was proposed to improve the classification accuracies of rice samples in LIBS technology. The classification accuracy of rice samples was increased by 20.75%. This method also achieved good results in the plastics, steel, rocks and minerals classification, with accuracies of 99.87, 99.44, 97.47, 91.25%, respectively. Compared to the traditional classification method such as FSIM, whose classification features could only be extracted from one-dimensional spectral intensities. However, IHFM could indirectly extract classification features from multi-dimensional spectral intensities by LIBS image. The features extraction way of the IHFM could effectively reduce the correlation between the characteristic matrices of the samples, so that the classification accuracies could be improved. IHFM provided an effective solution to the problem of qualitative analysis based on LIBS technology. The results demonstrated that the combination of LIBS technology with image processing technology and pattern recognition technology may promote LIBS analytical performance.