Open Access
Add to BibSonomy
Abstract
The whole ecosystem is suffering from serious plastic pollution. Automatic and accurate classification is an essential process in plastic effective recycle. In this work, we proposed an accurate approach for plastics classification using a residual network based on laser-induced breakdown spectroscopy (LIBS). To increasing efficiency, the LIBS spectral data were compressed by peak searching algorithm based on continuous wavelet, then were transformed to characteristic images for training and validation of the residual network. Acrylonitrile butadiene styrene (ABS), polyamide (PA), polymethyl methacrylate (PMMA), and polyvinyl chloride (PVC) from 13 manufacturers were used. The accuracy of the proposed method in few-shot learning was evaluated. The results show that when the number of training image data was 1, the verification accuracy of classification by plastic type under residual network still kept 100%, which was much higher than conventional classification algorithms (BP, kNN and SVM). Furthermore, the training and testing data were separated from different manufacturers to evaluate the anti-interference properties of the proposed method from various additives in plastics, where 73.34% accuracy was obtained. To demonstrate the superiority of classification accuracy in the proposed method, all the evaluations were also implemented by using conventional classification algorithm (kNN, BP, SVM algorithm). The results confirmed that the residual network has a significantly higher accuracy in plastics classification and shows great potential in plastic recycle industries for pollution mitigation.
© 2021 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Plastics are widely used in all walks of life with their excellent performance and rich diversity. They are not only in the manufacturing industry as the core materials such as anti-electromagnetic, extreme compression, but also an essential packaging material in our life, which bring a lot of convenience to people's life. The production of plastics has increased significantly over the past 60 years, where a large proportion of plastics had been used to make disposable packaging materials or other short-term products. Currently, these non-degradable plastics in our ecosystem have caused a large number of environmental problems, even threatening many creatures with extinction [1]. Classification of plastics is necessary to determine the disposal methods of different plastics after recycling. Conventional analytical techniques have been widely used in plastic recycling, such as chromatography, thermal analysis, infrared spectroscopy, ultraviolet-visible absorption spectrum, Raman spectrum, mass spectrometry and so on. Spectral analysis technology has the advantages of high sensitivity and so on [2–4]. However, most of these methods require complex and diverse sample pretreatment, lengthy analytical time, high-cost devices/materials, and high-level operation environment. Thus, they are generally operated in the laboratory, instead of industrial fields [5]. In past decades, scientists keep developing more convenient and efficient techniques to sort plastics.
Laser induced breakdown spectroscopy (LIBS) has highlighted its unique advantages in elemental analysis and shown a great application prospect. The principle of LIBS is to focus the high-energy laser pulse on the surface of the sample to generate plasma. In the plasma, the atoms, ions, and molecules in excited states transit down to lower states, emitting photons with a specific frequency and generating characteristic spectral lines. Then, the composition determination and classification can be carried out by analyzing the plasma spectrum [6]. Compared with the conventional methods, LIBS has the advantages of simultaneous analysis of multiple elements, no need for sample preparation, long-distance measurement, micro-destructive testing, and strong adaptability to the environment [7], which is in line with the current trend of science and technology in pursuit of efficient and green. LIBS shows its strong development potential in industrial manufacturing [8], environmental protection [9], biomedical [10], food [11], agriculture [12] and other fields.
Researchers have been exploring the application of LIBS in plastic classification, and some achievements have been made. The output of plasma spectral data of 15 plastic bonded explosives (PBXs) obtained by the LIBS analysis was in good agreement with the classification results obtained by dynamic mechanics analysis (DMA) [13]. The elemental and molecular information obtained from LIBS was effective in the classification of plastics, and LIBS technology had good spectral characteristic differences for colored plastic samples [14]. Y. Yu et al. [15] proposed to combine the adjusting spectral weighting (ASW) algorithm with LIBS, and then used Support Vector Machine (SVM) algorithm to classify the commonly used 11 kinds of plastics. The result of classification accuracy reached 100%. X.T. Yan et al. [16] used PCA and kNN algorithms for LIBS plastic classification. PCA was used for dimension reduction of data, and the average classification accuracy of 20 kinds of plastics under 20 dimensions was 99.6%. M. Boueri et al. [17] used LIBS technology to identify 8 kinds of polymer materials with artificial neural network (ANN), and the classification accuracy of each plastic was between 81%-100%. These studies demonstrated the feasibility to fast classify plastic material using LIBS technique with classification algorithms. Common classification algorithms include support vector machine (SVM) [18], k-nearest neighbor classification algorithm (kNN) [19], ANN (eg: back propagation (BP)) [20]. However, there is still some limitations and deficiency in these algorithms reported in previous works: SVM is too sensitive and dependent on parameter and kernel function selection, and it is difficult to implement for large-scale training; kNN would suffer from too large computation with large spectral data; ANN can hardly avoid over-learning and negative learning, and greatly different results in multiple training of the same data leads to poor prediction stability.
Deep learning, a branch of machine learning, has achieved excellent results in machine vision, speech recognition, natural language processing. In addition, increasing depth in deep learning networks can improve network performance, but excessive increase in depth will cause problems such as gradient dispersion and gradient explosion, which will lead to network performance degradation and achieve the counterproductive classification effect [21]. Residual network algorithm can solve these problems, and its internal residual block uses jump connection, which can alleviate the gradient disappearance and gradient explosion problems caused by increasing the network depth [22]. Therefore, the residual network is probably to improve the classification accuracy by adding considerable depth. However, there is no reported research on residual network in LIBS classification, to our best knowledge.
Generally, reducing training data results in accuracy deterioration. However, the training samples in industrial applications are usually limited, even only one sample available. It is necessary to establish a classification method with high performance in few-shot learning for industries. Moreover, the training set and test set in the classification data set were generally from the same samples in previous works [16,23], lack of anti-interference in industries. For example, there is a lack of test robustness research on plastics of the same kind but from another manufacturer. Various additives are generally mixed in plastic materials to realize specific performance, where the additive formulas are quite different in different manufacturers, even different production batches. These additives should not be the effective information in plastic classification. Therefore, for plastics of the same kind but with different contents of additives from different manufacturers, they should be classified as the same plastic in industrial recycling, which cannot be attributed to two kinds of plastics due to the interference of elements in additives.
In this paper, we proposed to accurately classify plastic materials by using residual network based on LIBS. Firstly, wavelet peak searching method was used to reduce the amount of spectral data. Then converting the original spectral images into the LIBS characteristic spectral images as the input image set of the residual network, training, verifying and testing of the residual network. Furthermore, few-shot learning in residual network was investigated. To demonstrate the anti-interference of various additives in plastics to classification results, the training and testing data were from different manufacturers (same main compositions with different additives such as plasticizers). The comparison of SVM, kNN, and ANN was also discussed in detail.
2. Experimental procedure and method
2.1 Experimental setup and sample
The structure of the experimental setup for LIBS used in this study is shown in Fig. 1. The experimental apparatus was mainly composed of laser, spectral acquisition system, spectrometer, intensified charge coupled device (ICCD) and spectral processing calculator. A Q-switched Nd: YAG laser (pulse width 6 ns, single pulse energy 40 mJ, wavelength of 532 nm after frequency doubling, the pulse repetition rate 10 Hz) was used. After passing through a system consisting of a reflector and a focusing lens (f=150 mm), the laser pulse ablated the sample surface to produce plasma. The sample was placed moving on an X-Y motorized stage. The optical emission of the plasma was collected coupled it to an optical fiber with a length of 2 m and a core diameter of 50 µm, then transmitted to the spectrometer (Andor ME 5000, λ/Δλ=5000, range 200-980 nm) and ICCD (Andor iStar 334 T). The ICCD started to record optical signal after 2.5 µs delay after laser ablation, and integrated for 2 µs. Each spectrum was accumulated by ten laser shots. The spectral data were transmitted to computer, and processed by classification tools.
In this work, acrylonitrile butadiene styrene (ABS), polyamide (PA), polymethyl methacrylate (PMMA), and polyvinyl chloride (PVC) from 13 manufacturers were used. All of these plastics were bulks and required no sample pretreatment. The total number of samples used in this experiment was 260. For each kind of samples, the numbers of blocks and spectra were 20 and 100, respectively.
2.2 Calculation process based on the residual network algorithm
For a shallow network with saturated accuracy, several identity mapping layers with output equal to input are added behind it to gain performance benefits by increasing the depth of network [24]. The higher accuracy would be obtained by continuous increasement of the layer number in the network. In the residual network, jump connection is added on the basis of the ordinary network, and a shortcut connection (also known as identity mapping) is added between each two layers of the network to form a residual block. The more residual blocks are set, the deeper the network depth is. According to the objects that are classified, the number of residual blocks should be selected on the basis of specific requirements. In Fig. 2, the residual network was formed by five residual block connections.
The specific structure of the residual block in a two-layer neural network is shown in Fig. 3. Assuming that ${a^{[l]}}$ is obtained through the previous network, it will serve as the input of the residual block. This can get ${a^{[l + 1]}}$ on the first activation, and then ${a^{[l + 2]}}$ on the second activation. The activation layer is first linearly activated to obtain
where ${W^{[l + 1]}}$ is the weight matrix, and ${b^{[l + 1]}}$ is the deviation factor. ${a^{[l + 1]}}$ can be obtained through Rectified Linear Units (ReLU) nonlinear activation function (g function):After going to the second level of activation and doing the linear activation again:
According to Eq. (3), the nonlinear activation of ReLU is repeated to get the expression for ${a^{[l + 2]}}$:
The above calculation process is the main line operation in the residual block, which shows how the data is transferred between two layers of the network in the residual block. This is also the transmission process of the data stream in the ordinary network. On this basis, the residual network also introduces a shortcut operation, which directly introduces the input data ${a^{[l]}}$ into the second activation layer of the residual block (after linear activation and before nonlinear activation of ReLU), so the expression of the ${a^{[l + 2]}}$ in Eq. (4) is modified as:
Combining Eqs. (1), (2), (3) and (5):
Equation (6) describes the relationship between input and output after data passing through two transmission channels, the main line and the shortcut simultaneously. In the process of training the deep network in our work, the training model was set to be an identity mapping. The input and output were exactly the same when passing through the network. In Eq. (6), the weight matrix ${W^{[l + 2]}}$ and deviation factor ${b^{[l + 2]}}$ are guaranteed to be 0, which can satisfy the requirement that the output after the two-layer neural network is equal to the original input data. Compared with the problem that it is difficult to set parameters when the identity function ${a^{[l + 2]}} = {a^{[l]}}$ is directly learned in the deep ordinary network, the residual network only needs to learn the identity function ${W^{[l + 2]}} = 0$ and ${b^{[l + 2]}} = 0$, which is simpler and more operable. In the residual block, the addition of the data on the main line and the shortcut not only retains the forward feature, but also prevents some problems such as gradient disappearance caused by the increase of network depth. Therefore, higher accuracy in residual network can be obtained in the deep network. The computer used for ResNet computing was Intel Core i5 8250U (4 cores, 8 threads, 1.6 GHz, 6 MB three-level cache), and the RAM (DDR4) was 8 GB and 2666 MHz. The motherboard chip of the computer was Dell 0HTGXJ (7th / 8th Generation Intel Processor Family I / O - 9D4E notebook chipset).
3. Results and discussion
3.1 Original LIBS spectra
Figure 4 shows the original LIBS spectra of four plastics from 13 different manufacturers. Moreover, the elements corresponding to the peak of the spectra were marked in Fig. 5. The major elements in the plastic are C, H, O and N (the corresponding spectral lines are C (247.9 nm), C-N (388.3 nm), C-C (516.5 nm), H (656.3 nm), N (742.4, 744.6, 746.8, 821.6, 844.7, 868.3 nm), O (777.3 nm)). In addition, there were spectral lines of minor elements (Mg (277.8, 279.1 nm), Si (288.2 nm), Ca (393.4, 396.8, 422.7 nm), Na (589.0, 589.6 nm)) provided by plastic additives [25]. Comparing the original spectra of the same plastic produced by different manufacturers, the differences in the spectra can be attributed to the different types and contents of additive elements. By comparing the original LIBS spectra of ABS and PA from different manufacturers, large differences were observed even if between the same species of plastics. In contrast, a smaller difference was observed between ABS1 and PA1. The spectral intensity reached its peak at the wavelength of about 400 nm, and the spectral intensity signals from 400 nm to 600 nm were weak, and small peaks of spectral intensity appeared after 600 nm. Plastics can be roughly classified by comparing the original LIBS spectrum of different plastics, but the classification became more difficult and unreliable when facing more types of samples. Therefore, we need some classification algorithms to obtain more accurate sample classification simply and quickly. In this study, the original LIBS spectrum of plastics was combined with residual network algorithm to carry out the accuracy classification of plastics.
3.2 Wavelet peak searching
In the processing of spectral data of plastic samples, there was a large amount of original spectral data, and most of the data actually made little contribution to classification, while some characteristic spectral lines made a greater contribution [26]. Continuous wavelet is an effective time-frequency analysis method, which can construct time-frequency representation of signals and provide time-frequency localization [27]. In this work, continuous wavelet peak searching was used to reduce the amount of data processing on the premise of less impact on the accuracy, and greatly improve the work efficiency. The concrete principle and method of continuous wavelet transform are as follows: after the continuous wavelet transform of the original spectral signal, the two-dimensional wavelet coefficient matrix was obtained. The higher the wavelet coefficient value is, the higher the similarity between the original spectral data and the selected fitting wavelet function is. Controlling the signal-to-noise ratio, the spectral peak was effectively searched and filtered. The scale function in the wavelet peak searching was an exponential base 1.18. The spectral intensities were transferred to 1024 pixels, corresponding to input images of 32*32*1 gray-scale maps in the residual network. SNR was controlled to make the peak points near 1024 after peak searching. Figure 6 shows the comparison of spectrum before and after peak searching. The number of peaks in the spectrum was greatly reduced after peak searching, while the effective peak points with high spectral intensity were all retained.
3.3 Evaluation of few-shot learning
First of all, the LIBS spectra of the plastic samples after peak searching were transformed to characteristic images (as shown in Fig. 7). The specific conversion process was to write the spectral intensity of each spectral pixel into the 8-bit grayscale image PNG format data structure in sequence from left to right and top to bottom. The black pixel indicated strong spectral intensity, while the white pixel indicated weak spectral intensity [28].
In this work, the plastic input image set was prepared according to CIFAR-10, an input image set of the residual network. CIFAR-10 consists of 60,000 32*32*3 color images from 10 classes. Each class contains 6,000 images, 5,000 training images and 1,000 test images. In this plastic classification based on residual network, spectra of each plastic sample were transformed to 100 images. A total of 1300 transformed images (1024 pixels) were generated as the input of the residual network. All these data were separated randomly into two halves of training set and verification set. The learning rate and the network depth were set to be 0.01 and 14 respectively. Further increasing the depth resulted in little accuracy improvement but much longer classification time. Under the above parameters, the training process and results are shown in Fig. 8. After 15 rounds of training (150 iterations), the accuracy of training and verification both reached 100%. The specific classification of each plastic sample is shown in Fig. 9. The results indicated that all the verification data were correctly classified.
To demonstrate the ability of residual network algorithm in few-shot learning, the number of training images for each sample was reduced. The comparison with other algorithms of BP, kNN, and SVM was also made under the same conditions. The results were shown in Table 1. When the training number was greater than 5, the verification accuracy of classifying plastics of the same type and different manufacturers into one category by the four classification algorithms almost reached 100%. However, the accuracy of kNN and SVM algorithm decreased sharply (30.77% and 64.10%) under one spectrum training. By contrast, the residual network kept 100% accuracy of verification set, which was also higher than 81.74% of BP algorithm. Therefore, residual network can play its role with limited data for training in industrial fields, even with only one training data.
Table 1. Validation results of residual network, BP, kNN and SVM algorithm after few-shot learning
3.4 Evaluation of anti-interference
To demonstrate the anti-interference of various additives in plastics to classification results, each type of samples from one of the manufacturers were used as the training set, while samples from other manufacturers were the testing set. Each plastic in the training set had 80 LIBS characteristic spectra, and each plastic in the test set had 20 LIBS characteristic spectra. The learning rate and the network depth were 0.001 and 14 respectively. The classification was repeatedly implemented five times under the same conditions. With the training network of 14 layers, the average classification accuracy of 73.34% of the test set was obtained.
Similarly, the same training set and test set were used as the input of BP, kNN and SVM algorithms. And the classification accuracy of test set of each algorithm was shown in Fig. 10 below. In the kNN classification algorithm, the value of k was selected to be 7, and the average classification accuracy of 9 untrained plastic test sets was 66.67%, and the plastics of three manufacturers were wrongly classified. And the average classification accuracy of 9 test samples in SVM classification algorithm was 55.00%, which the optimal parameters were selected in the range [-10, 10] for SVM network training. For BP classification algorithm, the classification accuracy of the test set was very unstable. Under the same test conditions, it could span more than 20% and its change was irregular, therefore 100 times classification was repeatedly implemented to make average. Under different network depths, the average test accuracy was different, and the highest test average classification accuracy was 57.91% in the 1-25 layer network. The results demonstrated higher accuracy in residual network than those in kNN, SVM and BP algorithms. Moreover, the criterion and method to determine the number of hidden layers and nodes in BP network are unclear at present. The conventional method needs to set and adjust the parameters continuously and determines the final number of hidden layers and nodes according to the optimal error results of network. It is time-consuming and laborious. The setting of BP network needs a long time of manual adjustment. In residual network, the important parameters are the network depth and the number of iterations of each training. The setting of these parameters has a certain setting rules. For example, when the residual block is 3 layers, the network depth can only be multiple of three plus two layers (initial input layer and final output layer). There is no need to adjust the number of neural nodes after setting the network depth like BP. Both as neural networks, residual network is much more convenient in parameter setting than BP.
The following is the reason why the anti-interference performance of the residual network will be better than other traditional classification algorithms, such as BP. There are two main algorithms based on ANN: one is to extract the features of the image to be recognized, and then use the obtained features to train the neural network classifier; the other one is to directly input the image to be processed into the network, and then the network automatically extracts the features until the recognition result is obtained. The input data format of BP network is required to be a vector, so the BP feature extraction based on spectra is one-dimensional. The residual network extracts features according to the unit region on the image, and obtains multi-dimensional features through the gray gradient relationship of all adjacent pixels in the region. The uncertainty of signal will affect the accuracy of multivariable model, and the accuracy can be improved by improving the repeatability of signal to reduce the uncertainty of signal [29]. Multi-dimensional data features were used in residual network, while only one-dimensional data features were used in conventional neural network algorithms, such as BP. Therefore, the utilization of signal data in the residual network has high repeatability and makes its classification performance better.
Though the results show great superiority in residual network algorithm compared with conventional methods, further improvement is still needed in our future work. Continued optimization of activation function selection, pre-training, and GPU parallel operation in residual network algorithm may have some positive impact.
4. Conclusion
In this work, we proposed an accurate method for plastics classification based on LIBS and residual network algorithm for plastic recycle industries. The LIBS spectral data were compressed by peak searching algorithm based on continuous wavelet; then compressed data were transformed to characteristic image for residual network input to make training, verification and testing. When the training image data of each kind of plastic was in the extreme condition of only one image, the classification accuracy of this method in few-shot learning still kept 100%, which was much higher than the conventional classification algorithms (kNN, SVM, and BP). Moreover, anti-interference properties from various additives in plastics to classification accuracy was firstly evaluated to separate training and testing sample from different manufacturers. The result in the residual network algorithm was 73.34%, significantly higher than 66.67%, 55.00%, and 57.91% accuracy in kNN, SVM, and BP, respectively. Further improvement is still needed in our future work to increase reliability. This work demonstrated combination of LIBS and residual network was an effective and accurate approach for plastic classification, and shows great prospect in the field of plastic pollution mitigation.
Funding
Key-Area Research and Development Program of Guangdong Province (2020B090922006); National Natural Science Foundation of China (62005081, 62105105); Basic and Applied Basic Research Foundation of Guangdong Province (2019A1515111120, 2020A1515110985, 2021A1515011932); Guangzhou Municipal Science and Technology Project (202002030165); Featured Innovation Project of Guangdong Education Department (2019KTSCX034); Young Scholar Foundation of South China Normal University (19KJ13); Special Funds for the Cultivation of Guangdong College Students' Scientific and Technological Innovation ("Climbing Program" Special Funds) (pdjh2020b0153).
Disclosures
The authors declare that there are no conflicts of interest related to this paper.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. J. Hopewell, R. Dvorak, and E. Kosior, “Plastics recycling: challenges and opportunities,” Philos. Trans. R. Soc., B 364(1526), 2115–2126 (2009). [CrossRef]
2. Z. Lang, S. Qiao, Y. He, and Y. Ma, “Quartz tuning fork-based demodulation of an acoustic signal induced by photo-thermo-elastic energy conversion - ScienceDirect,” Photoacoustics 22, 100272 (2021). [CrossRef]
3. Y. Ma, Y. Hu, S. Qiao, Y. He, and F. K. Tittel, “Trace gas sensing based on multi-quartz-enhanced photothermal spectroscopy,” Photoacoustics 20, 100206 (2020). [CrossRef]
4. S. Qiao, Y. Ma, Y. He, P. Patimisco, A. Sampaolo, and V. Spagnolo, “Ppt level carbon monoxide detection based on light-induced thermoelastic spectroscopy exploring custom quartz tuning forks and a mid-infrared QCL,” Opt. Express 29(16), 25100–25108 (2021). [CrossRef]
5. Y. Yang, X. Zhang, J. Yin, and X. Yu, “Rapid and Nondestructive On-Site Classification Method for Consumer-Grade Plastics Based on Portable NIR Spectrometer and Machine Learning,” Journal of Spectroscopy 2020, 1–8 (2020). [CrossRef]
6. D. Kim, J.-H. Yang, S. Choi, and J. J. Yoh, “Analytical Methods to Distinguish the Positive and Negative Spectra of Mineral and Environmental Elements Using Deep Ablation Laser-Induced Breakdown Spectroscopy (LIBS),” Appl. Spectrosc. 72(6), 896–907 (2018). [CrossRef]
7. B. B. S. Jaswal, V. Kumar, J. Sharma, P. K. Rai, M. A. Gondal, B. Gondal, and V. K. Singh, “Analysis of heterogeneous gallstones using laser-induced breakdown spectroscopy (LIBS) and wavelength dispersive X-ray fluorescence (WD-XRF),” Lasers Med 31(1), 1–6 (2016). [CrossRef]
8. R. Noll, C. Fricke-Begemann, M. Brunk, S. Connemann, and C. Meinhardt, “Laser-induced breakdown spectroscopy expands into industrial applications,” Spectrochim. Acta, Part B 93, 41–51 (2014). [CrossRef]
9. Q. Zhang, Y. Liu, Y. Chen, Y. Zhangcheng, and L. Li, “Online-Detection of Halogen Atoms in Atmospheric VOCs by LIBS-SPAMS Technique,” Opt. Express 28(15), 22844–22855 (2020). [CrossRef]
10. F. Mehari, M. Rohde, C. Knipfer, R. Kanawade, F. Klämpfl, W. Adler, F. Stelzle, and M. Schmidt, “Laser induced breakdown spectroscopy for bone and cartilage differentiation - ex vivo study as a prospect for a laser surgery feedback mechanism,” Biomed. Opt. Express 5(11), 4013–4023 (2014). [CrossRef]
11. P. Yang, X. Li, and Z. Nie, “Determination of nutrient profile in plant materials using laser-induced breakdown spectroscopy with partial least squares-artificial neural network hybrid model,” Opt. Express 28(15), 23037–23047 (2020). [CrossRef]
12. J. N. Kunz, D. V. Voronine, H. W. H. Lee, A. V. Sokolov, and M. O. Scully, “Rapid detection of drought stress in plants using femtosecond laser-induced breakdown spectroscopy,” Opt. Express 25(7), 7251 (2017). [CrossRef]
13. S. H. Ahmadi, M. H. Keshavarz, and H. R. H. Atabak, “Correlations Between Laser Induced Breakdown Spectroscopy (LIBS) and Dynamical Mechanical Analysis (DMA) for Assessment of Aging Effect on Plastic Bonded Explosives (PBX),” Z. Anorg. Allg. Chem. 645(2), 120–125 (2019). [CrossRef]
14. K. M. M. Shameem, K. S. Choudhari, A. Bankapur, S. D. Kulkarni, V. K. Unnikrishnan, S. D. George, V. B. Kartha, and C. Santhosh, “A hybrid LIBS–Raman system combined with chemometrics: an efficient tool for plastic identification and sorting,” Anal. Bioanal. Chem. 409(1), 1–2 (2017). [CrossRef]
15. Y. Yu, L. B. Guo, Z. Q. Hao, X. Y. Li, M. Shen, Q. D. Zeng, K. H. Li, X. Y. Zeng, Y. F. Lu, and Z. Ren, “Accuracy improvement on polymer identification using laser-induced breakdown spectroscopy with adjusting spectral weightings,” Opt. Express 22(4), 3895 (2014). [CrossRef]
16. X. Yan, X. Peng, Y. Qin, Z. Xu, B. Xu, C. Li, N. Zhao, J. Li, Q. Ma, and Q. Zhang, “Classification of plastics using laser-induced breakdown spectroscopy combined with principal component analysis and K nearest neighbor algorithm - ScienceDirect,” Results in Optics 4, 100093 (2021). [CrossRef]
17. M. Boueri, V. Motto-Ros, W. Q. Lei, Q. L. Ma, and Y. Jin, “Identification of Polymer Materials Using Laser-Induced Breakdown Spectroscopy Combined with Artificial Neural Networks,” Appl. Spectrosc. 65(3), 307–314 (2011). [CrossRef]
18. T. Zhang, S. Wu, J. Dong, J. Wei, K. Wang, H. Tang, X. Yang, and H. Li, “Quantitative and classification analysis of slag samples by laser induced breakdown spectroscopy (LIBS) coupled with support vector machine (SVM) and partial least square (PLS) methods,” J. Anal. At. Spectrom. 30(2), 368–374 (2015). [CrossRef]
19. A. M. Neiva, M. A. C. Jacinto, M. M. d. Alencar, S. N. Esteves, and E. R. Pereira-Filho, “Proposition of classification models for the direct evaluation of the quality of cattle and sheep leathers using laser-induced breakdown spectroscopy (LIBS) analysis,” RSC Adv. 6(106), 104827 (2016). [CrossRef]
20. Z. Lu, J. Mo, S. Yao, J. Zhao, and J. Lu, “Rapid Determination of the Gross Calorific Value of Coal Using Laser-Induced Breakdown Spectroscopy Coupled with Artificial Neural Networks and Genetic Algorithm,” Energy Fuels 31(4), 3849–3855 (2017). [CrossRef]
21. L. Liang, J. Cao, X. Li, and J. You, “Improvement of Residual Attention Network for Image Classification,” Springer, Cham 11935, 529–539 (2019). [CrossRef]
22. K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” IEEE 1, 770–778 (2016). [CrossRef]
23. K. Liu, C. Qiu, D. Tian, G. Yang, Y. Li, and X. Han, “Laser-Induced Breakdown Spectroscopy for Plastic Classification,” Spectroscopy and Spectral Analysis 37, 3600–3605 (2017).
24. L. Yu, H. Chen, Q. Dou, J. Qin, and P. A. Heng, “Automated Melanoma Recognition in Dermoscopy Images via Very Deep Residual Networks,” IEEE Trans. Med. Imaging 36(4), 994–1004 (2017). [CrossRef]
25. Q. Wang, X. Cui, G. Teng, Y. Zhao, and K. Wei, “Evaluation and improvement of model robustness for plastics samples classification by laser-induced breakdown spectroscopy,” Opt. Laser Technol. 125, 106035 (2020). [CrossRef]
26. Y. Zhang, C. Sun, Z. Yue, S. Shabbir, and J. Yu, “Correlation-based carbon determination in steel without explicitly involving carbon-related emission lines in a LIBS spectrum,” Opt. Express 28(21), 32019 (2020). [CrossRef]
27. Y. Zheng, D. Tian, K. Liu, Z. Bao, P. Wang, C. Qiu, D. Liu, and R. Fan, “Peak Detection of TOF-SIMS using Continuous Wavelet Transform and Curve Fitting,” Int. J. Mass Spectrom. 428, 43–48 (2018). [CrossRef]
28. J. Yan, P. Yang, Z. Hao, R. Zhou, X. Li, S. Tang, Y. Tang, X. Zeng, and Y. Lu, “Classification accuracy improvement of laser-induced breakdown spectroscopy based on histogram of oriented gradients features of spectral images,” Opt. Express 26(22), 28996–29004 (2018). [CrossRef]
29. Z. Wang, M. S. Afgan, W. Gu, Y. Song, and Z. Li, “Recent Advances in Laser-induced Breakdown Spectroscopy Quantification: from Fundamental Understanding to Data Processing,” TrAC Trends in Analytical Chemistry 143, 116385 (2021). [CrossRef]
