Classification of Metal and Metal Oxide Nanoparticles Using Machine Learning and Deep Learning Models

doi:10.21203/rs.3.rs-3201733/v1

For decades classification of objects based on texture has portrayed a prominent role in the applications of computer vision and image processing. The need for automated classification of objects sharing similar visual appearance is growing day by day in various industries like metal, paper, wood, etc. The initial task of any metal devising industry is to classify the metals before they are used in building any devices. For ages, traditional approaches are used to identify the metals using their properties which is time-consuming and not economical. In the proposed study an automated model for texture analysis and classification of the metal and metal oxide nanoparticles is developed using machine learning and deep learning concepts to overcome the drawbacks of the traditional approach. The machine learning model uses KNN and PNN classifiers, and the deep learning model uses LeNet, and ConvXGB classifiers to analyze the texture and classify them as metals (silver, boron) and metal oxides (iron oxide, copper oxide). From the experiment it is found that the average accuracy using the KNN classifier is 70.00%, the PNN classifier is 75.00%, LeNet is 95.00%, and ConvXGB yields 85.00%. It is analyzed that the LeNet has the highest accuracy of 95%, and hence, it is suitable for the classification of silver, boron, iron oxide, and copper oxide nanoparticle images.

Metals

Metal oxides

Texture analysis

Machine learning

Deep learning

Classification

KNN

PNN

LeNet

ConvXGB

Today nanotechnology is being used in most disciplines. In recent decades, the combination of artificial intelligence and nanotechnology has made a buzz in the research world also. Machine learning and deep learning models have been widely used to achieve better results in the areas like computer vision, big data, automatic voice recognition, and natural language processing. These models have also shown their prominence in metal device engineering industries. The metals need to be detected and their properties have to be analyzed to match the device to be engineered. The traditional approach used to identify the metals in metal device construction industries is time-consuming, requires more manpower, and is not cost-effective. In the proposed study an automated method is developed by extracting texture features and classification of the metal and metal oxide is done using machine learning and deep learning models [1, 2] to overcome the drawbacks of the traditional approach.

Over the last few decades, deep learning tools are being accepted widely due to their ability to handle large amounts of data. Convolutional Neural Network (CNN) is one of the most popular neural networks in deep learning. CNNs represent a significant advancement in pattern recognition, voice recognition, natural language processing, and video processing. Deep learning approaches such as Convolutional Neural Networks (CNN) [3] have recently shown excellent image classification performance over machine learning classification. CNNs can learn rich features from images, as opposed to hand-crafted features. These networks, which have been rigorously trained on a large number of images, can be used as feature extractors in a variety of computer vision scenarios. CNN's capacity to learn complicated features makes it more successful in detecting texture features of nanoparticle images. Figure 1, shows the general architecture of the CNN model [4].

CNN layers begin by learning features such as lines, edges, and textures. The next layers discover complex patterns and textures, while the last layers discover the critical features required for application categorization

In the present study, statistical texture analysis [5] is performed by extracting the texture features such as kurtosis, skewness, and entropy from the nanoparticle images FESEM/TEM to identify the metals and the metal oxides nanoparticle using machine learning and deep learning models. The machine learning model is constructed using the KNN [6], and PNN [7] classifiers, and the deep learning model is constructed using LeNet, and CONVXGB classifiers.

Many researchers have contributed their work using machine and deep learning concepts in various areas. Sultana et. al [8] have explained various advancements in image classification using Convolutional Neural Networks. Zhijian Sun et al. [9] have suggested a system that consists of three stages: nanoparticle shape extraction, nanoparticle segmentation using a powerful lightweight deep learning network (NSNet), and statistical analysis on SEM and TEM images. M. Keyed et. al [10] have proposed CNN-based LeNet-5 architecture to train parameters of the CNN on the Fashion MNIST dataset for clothes classification. Lijie Zhou et. al [11] have proposed an improved convolution neural network combined with LBP on handwritten digits and face datasets for face recognition research. R. A. Hazarika et. al [12] have performed the classification of Alzheimer’s Disease on brain MRI images using LeNet-Deep Neural Network Model. Shuai Wang et. al [13] used Deep CNN in detecting the defects on steel surfaces. Deep learning was used by Tuan Linh Dang et. al [14] in detecting the metals in the debris. Texture directionality detection was carried out by Kociołek, M [15], using Deep learning models. Eftekhar Hossain et. al [16] used the KNN approach to detect and classify the diseases in leaves. Turgut [17] classified infant cries using different deep neural network models. Nano polymer surface analysis using a machine learning algorithm was proposed by Małgorzata S. et. al [18]. The roughness of the metal was studied by Pan Z. et. al [19] using ML classifiers. Díaz-Romeroet. Al [20] suggested a mass prediction system on a combined Cast (C), Wrought (W), and Stainless Steel (SS) image dataset using machine learning and deep learning algorithms.

The experimentation is done on nanoparticle images of boron, iron oxide, silver, and copper oxide. The metal and metal oxide nanoparticle images used in the experiment underwent different synthesization processes and the characterization is performed in various institutions like IISC including Kachin, Bangalore, and Mysore. The synthesis process is carried out in the Department of Chemistry, Rani Channamma University, Belagavi. The details of the synthesis process used for various metal and metal oxides is as follows;

A. Boron FESEM nanoparticle images

Boron FESEM nanoparticle images were obtained from a chemically synthesized process, where a combination of diborane (B2H6) and nitrogen (N2) is driven into a chamber where a voltage of 20 to 30 volts is placed between the cathode and anode for evaporation. The deposits on the chamber's water-cooled wall were collected. The deposits were formed on an aluminum oxide crucible and then heated for 4 hours at 1073 K. Finally, the sample was cooled to room temperature using just N2. Finally, the samples were sent for characterization to obtain the Boron FESEM nanoparticle images. Figure 2(a) shows the sample images of Boron nanoparticles used for training and testing.

B. Iron oxide TEM nanoparticle images

The dataset in the study is made up of multiple TEM nanoparticle images of Iron Oxide that were synthesized and generated utilizing two separate synthesis processes, namely chemical and biological. Table 1 lists the details of chemicals used in obtaining the Iron oxide TEM nanoparticles, as well as the temperature maintained during the production process. Figure 2(b) shows the sample nanoparticle images of Iron oxide.

Table 1

The details of chemicals used in obtaining the Iron oxide TEM nanoparticles
Image	Chemicals used in the synthesis of the nanoparticles	Temperature (⁰C)
A	Prepared by Iron Nitrate	300
B	Prepared by FeCl₂, 2H₂O	300
C	Prepared by Iron Nitrate	500
D	Prepared by FeCl₂, 2H₂O	500
E	Prepared by Iron Nitrate using plant extract	300

C. Silver FESEM nanoparticle images

Silver nanoparticles were created by altering the pH of Silver Nitrate Green synthesis procedures while maintaining other parameters such as concentration, duration, and temperature constant. Table 2. Shows details of chemical compositions for the preparation of Silver nanoparticles. Figure 2(c) depicts the sample images used in the experiment.

Table 2

The details of chemical compositions for the preparation of Silver nanoparticles
Image	Concentration (mM)	Time (min)	Temperature (0C)	pH
A	1	30	90	3
B	1	30	90	5
C	1	30	90	6
D	1	30	90	10

D. Copper oxide FESEM nanoparticle images

The pictures of Copper Oxide FESEM nanoparticles used in the proposed study were collected through the synthesis and characterization process. The precipitation technique is used in the synthesis, in which copper nitrate (Cu(NO₃)2H₂O) is dissolved in 100 ml deionized water to generate a 0.1 M concentration. The pH of the NaOH solution (0.1 M) was gradually increased while vigorously swirling until it reached 11. Black precipitates were produced and rinsed many times with deionized water and 100% ethanol until the pH reached 7. Following that, the washed precipitates were dried for 16 hours at 80°C. Finally, the precursors were calcined for 4 hours at 500°C to get the samples. Figure 2(d) displays the sample nanoparticle images of copper oxide FESEM images used in the experiment.

The proposed method is used to perform the texture analysis by constructing two models; machine learning and deep learning. These models emphasize extracting the texture features of the nanoparticles in the input images of the metal (silver, boron) and metal oxides (iron oxide, copper oxide) and then classifying them based on their texture features. KNN and PNN classifiers are used in machine learning models and deep learning models incorporate two advanced classifiers namely LeNet and ConvXGB.

The algorithm for the machine learning model is as follows;

Algorithm 1

Training phase

Step 1: Read FESEM and TEM images of boron, iron oxide, silver, and copper oxide nanoparticle

from the training set.

Step 2: Convert the given color image into a grayscale.

RGB values converted grayscale values by forming a weighted sum of the R, G, and B

components: gray = 0.2989 * R + 0.5870 * G + 0.1140 * B.

Step 3: Binarize the image using global thresholding to get the mask.

Step 4: Remove the non-pore part using a mask.

Step 5: Calculate the mean, kurtosis(k), skewness(s), and entropy(e) which has been

described in Eqs. (1), (2), and (3).

Step 6: Accumulate the extracted features in the vector form as a knowledge base.

Step 7: Stop.

Algorithm 2

Testing phase

Step 1: Read FESEM and TEM images of boron, silver, and copper oxide nanoparticle

from the testing set.

Step 2: Convert the given color image into a grayscale.

RGB values converted grayscale values by forming a weighted sum of the R, G, and B

components: gray = 0.2989 * R + 0.5870 * G + 0.1140 * B.

Step 3: Binarize the image using global thresholding to get the mask.

Step 4: Remove the non-pore part using a mask.

Step 5: Calculate the mean, kurtosis(k), skewness(s), and entropy(e).

Step 6: Performed classification using PNN and K-NN classifiers for classifying the images as Boron,

Iron oxide, Silver, and Copper oxide.

Step 7: Finally, calculate performance analysis parameters for measuring the quality

of the nanoparticles.

Step 8: Stop.

The flow diagram of the proposed machine learning model is depicted in Fig. 3.

The algorithm for the deep learning model is as follows;

Algorithm 1

Training phase

Step 1: Input FESEM and TEM images of boron, iron oxide, silver, and copper oxide nanoparticle

from the training set.

Step 2: Scale the images to fit the model.

Step 3: Convert the image to an image array.

Step 4: Build the corresponding label array on step 3.

Step 5: Build a CNN model using LeNet and ConvXGB.

Step 7: Define data augmentation.

Step 8: Fit the training models.

Step 9: Store the model in a knowledge database.

Step 10: Stop.

Algorithm 2

Testing phase

Step 1: Input FESEM and TEM images of boron, iron oxide, silver, and copper oxide nanoparticle

from the training set.

Step 2: Scale the images to fit the model.

Step 3: Convert the image to an image array.

Step 4: Build the corresponding label array on step 3.

Step 5: Build a CNN model using LeNet and ConvXGB.

Step 7: Define data augmentation.

Step 8: Fit the training models.

Step 9: Load the model from the training knowledge database.

Step 10: Predict the nanoparticle images as Boron, Iron oxide, Silver, and Copper oxide.

Step 11: Stop.

The flow diagram of the proposed deep learning mois del shown in Fig. 4.

The experimentation is done on 350 images of various metal and metal oxide nanoparticles and implemented on MATLAB R2018a, Python software on Intel(R) Core™ i5-10210U [email protected] system. The proposed method is accomplished using two models: machine learning and deep learning. The emphasis is on the texture analysis of the metal (silver, boron) and metal oxides (iron oxide, copper oxide). An automated tool is developed using machine learning models such as KNN, PNN, and deep learning models that incorporate two advanced classifiers; LeNet and CONVXGB. In the machine learning model, various textural features such as kurtosis, skewness, and entropy are extracted and classification is done by using KNN and PNN classifiers.

In deep learning, a batch size of 32 is processed to build the corresponding label array. Using this label array the CNN model (LeNet and ConvXGB) is constructed and then the augmentation is defined for the corresponding model and is the made fit for training. This model is stored for future testing. In testing, the stored model is loaded and the type of metal or metal oxide is predicted based on the features of the nanoparticles in the testing image.

The performance of the proposed algorithm in both the machine learning and deep learning models is evaluated based on performance evaluation measures such as; true positive (TP), true negative (TN), false positive (FP), false negative (FN), precision(Pr), recall(Re), and accuracy(Ac). Higher true positive values and accuracy show the effectiveness of the algorithm.

The statistical texture features of the metal and metal oxides used by the machine learning and deep learning models for classification are weighed using the following Equations (1), (2), and (3);

Kurtosis \(k=\sum \frac{{\left(Ii-\mu \right)}^{4}}{{n\sigma }^{4}}\) where \(\sigma\) – standard deviation (1)

Skewness \(s=\sum \frac{{\left(Ii-\mu \right)}^{3}}{{n\sigma }^{3}}\) (2)

Entropy e = sum (p.*log2 (p)) where p = normalized histogram counts (3)

4.1 Machine learning:

The proposed machine learning algorithm includes two phases, namely; the training phase (phase 1) and the testing phase (phase 2). In phase 1, the preprocessing techniques are adopted such as; the removal of unwanted background, converting images to gray images, binarising given images by applying global thresholding to get the mask, and calculate the mean of nanoparticles, and finally extracting the texture feature namely; kurtosis, skewness, and entropy, etc. Based on these texture features, the images are classified by using PNN and K-NN classifiers. In phase 2, the same preprocessing techniques are applied to the testing data set as used in phase 1, and the texture features are extracted from all the nanoparticle images. Finally, the performance evaluation of the nanoparticle images is carried out by calculating the accuracy, precision, and recall.

In the machine learning model, various textural features such as kurtosis, skewness, and entropy are computed, which are stored in the knowledge base. using KNN and PNN classifiers. The true positive (TP), true negative (TN), false positive(FP), false negative (FN), precision (Pr), recall (Re), and accuracy(Ac) are computed from the proposed classifiers and it is found that the KNN classifier has an accuracy of 85.00%, 85.00%, 80.00%, and 90.00% for silver, boron, iron oxide and copper oxide respectively with an average accuracy of 70.00%. The PNN classifier yields an accuracy of 85.00%, 85.00%, 95.00%, and 85.00% for silver, boron, iron oxide, and copper oxide respectively with an average accuracy of 75.00%.

4.2 Deep learning:

The suggested deep learning algorithm is divided into two phases: training (phase 1) and testing (phase 2). In the training phase, the preprocessing techniques such as; scaling the images, and converting the images as image arrays, followed by building the corresponding label array, building a CNN model using LeNet and ConvXGB classifiers, defining data augmentation, fitting the training model, and finally storing the extracted features such as; skewness, entropy, and kurtosis in a knowledge database is done. In the testing phase, similar preprocessing operations of the training phase are performed. Then similar texture features such as skewness, entropy, and kurtosis are extracted and stored as a knowledge database. These features are compared with the features stored as a knowledge base of the testing phase, and the prediction of the type of metal (silver, boron) and metal oxide (iron oxide, copper oxide) is accomplished. Finally, the performance evaluation of the algorithm can be computed by calculating the true positive, true negative, false positive, false negative, accuracy, precision, and recall.

Deep learning classification is performed to determine the robustness of the proposed system using the LeNet, and ConvXGB classifiers. The LeNet classifier provides an accuracy of 100%, 100%, 80.00%, and 100.00% for silver, boron, iron oxide, and copper oxide images respectively with an average accuracy of 95.00%. The ConvXGB classifier gives an accuracy of 80.00%, 80.00%, 100%, and 80.00% for silver, boron, iron oxide, and copper oxide nanoparticle images respectively, with an average accuracy of 75.00%. The details of classification results between machine learning and deep learning models are tabulated in Table 3 and the graph representing average accuracy using KNN, PNN, LeNet, and ConvXGB models is depicted in Fig. 5.

Table 3

The details of classification results between machine learning and deep learning models
Models	Classification method	Nanoparticle Images	Classification results
Models	Classification method	Nanoparticle Images	TP	FP	FN	TN	Pr	Re	Accuracy (%)
Machine learning	KNN	Boron	5	3	0	12	0.80	1.0	85.00
		Copper Oxide	4	1	1	14	0.93	0.80	90.00
		Iron Oxide	3	2	2	13	0.86	0.60	80.00
		Silver	2	0	3	15	1.0	0.40	85.00
	Avg. Accuracy 70.00%
	PNN	Boron	5	3	0	12	0.1	0.80	85.00
		Copper Oxide	4	2	1	13	0.86	0.80	85.00
		Iron Oxide	4	0	1	15	1.00	0.80	95.00
		Silver	2	0	3	15	1.00	0.40	85.00
	Avg. Accuracy 75.00%
Deep learning	LeNet	Boron	35	0	0	0	0.83	1.00	100
		Copper Oxide	35	0	0	0	1.00	1.00	100
		Iron Oxide	28	0	0	7	1.00	0.80	80.00
		Silver	35	0	0	0	1.00	1.00	100
	Avg. Accuracy 95.00%
	ConvXGB	Boron	28	0	7	0	0.80	0.80	80.00
		Copper Oxide	28	0	7	0	1.00	0.80	80.00
		Iron Oxide	35	0	0	0	0.71	1.00	100
		Silver	28	0	7	0	1.00	0.80	80.00
	Avg. Accuracy 85.00%

TP- True Positive, TN - True Negative, FP - False Positive, FN - False Negative, Pr – Precision, Re – Recall,

PNN –Probabilistic neural network, KNN-K-nearest neighbor

From the present study, it is analyzed that the LeNet has the highest accuracy of 95%, for the classification of metal and metal oxide nanoparticle images and hence, it is suitable for the classification of silver, boron, iron oxide, and copper oxide nanoparticle images.

In the proposed study an automated texture analysis and classification of the metals and metal oxide model is developed using machine learning and deep learning concepts to overcome the drawbacks of the traditional approach. The machine learning model uses KNN and PNN classifiers, and the deep learning model uses LeNet, and ConvXGB classifiers to analyze the texture of metals (silver, boron) and metal oxides (iron oxide, copper oxide). From the experiment it is observed that the average accuracy using the KNN classifier is 70.00%, the PNN classifier is 75.00%, LeNet is 95.00%, and ConvXGB yields 85.00%. It is analyzed that the LeNet has the highest accuracy of 95%, and hence, it is suitable for the classification of silver, boron, iron oxide, and copper oxide nanoparticle images.

Acknowledgments

The authors are grateful to the ‘KSTEPS, DST, GOVT. OF KARNATAKA’ for providing financial assistance and sanctioning a Ph.D. fellowship to carry out this research work. The authors are also grateful to the Department of Chemistry, Rani Channamma University, Belagavi for providing all the metals (boron, silver), and metal oxide (iron oxide, copper oxide) nanoparticle images and visualizing the manual results.

Wang, P., Wang, E.F.P.: Comparative analysis of image classification algorithms based on traditional machine learning and deep learning. Pattern Recognit. Lett. 141, 61–67 (2021). https://doi.org/10.1016/j.patrec.2020.07.042
Zhang, Dengsheng: Texture Feature Extraction, Fundamentals of Image Data Mining: Analysis, Features, Classification and Retrieval, pp-81-111. (2019)
Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: CNN features off-the-shelf: an astounding baseline for recognition, IEEE computer society conference on Computer Vision and Pattern Recognition, (2014). https://doi.org/10.48550/arXiv.1403.6382
Gu, H., Wang, Y., Hong, S.: Gui, Guan, Blind Channel Identification Aided Generalized Automatic Modulation Recognition Based on Deep Learning. IEEE Access. 1–1 (2019). 10.1109/ACCESS.2019.2934354
Ramola, A., Shakya, A.K., Van Pham, D.: Study of statistical methods for texture analysis and their modern evolutions. Eng. Rep. 2(4), 1–24 (2020). https://doi.org/10.1002/eng2.12149
https://machinelearningmastery.com/k-nearest-neighbors-for-machine-learning
Lavanyadevi, R., Machakowsalya, M., Nivethitha, J., Niranjil, A., Kumar: Brain tumor classification and segmentation in MRI images using PNN, IEEE International Conference on Electrical, Instrumentation, and Communication Engineering, Karur India 2017, pp.1–6, doi: 10.1109/ICEICE.2017.8191888
Sultana, F., Sufian, A., Dutta, P.: Advancements in Image Classification using Convolutional Neural Network, pp.122–129, (2019). https://doi.org/10.1109/ICRCICN.2018.8718718
Zhijian, S., Jia, S., Wang Jian, L.M., Wang Zhuo, B.X., Wang, Xiaoxiong, Nanoscale: Vol.14, pp.10761–10772, (2022). https://doi.org/10.1039/D2NR01029A
Kayed, M., Anter, A., Mohamed, H.: Classification of Garments from Fashion MNIST Dataset Using CNN LeNet-5 Architecture, International Conference on Innovative Trends in Communication and Computer Engineering (ITCE), Aswan, Egypt, 2020, pp. 238–243, doi: 10.1109/ITCE48509.2020.9047776
Lijie Zhou, W., Yu: Improved Convolutional Neural Image Recognition Algorithm based on LeNet-5, Journal of Computer Networks and Communications, Vol. 2022, (2022). https://doi.org/10.1155/2022/1636203
Hazarika, R.A., Abraham, A., Kandar, D., Maji, A.K.: An Improved LeNet-Deep Neural Network Model for Alzheimer’s Disease Classification Using Brain Magnetic Resonance Images, in IEEE Access, Vol. 9, pp. 161194–161207, doi: (2021). 10.1109/ACCESS.2021.3131741
Shuai Wang, X., Xia, L., Ye, B., Yang: Automatic Detection and Classification of Steel Surface Defect Using Deep Convolutional Neural Networks, Application of Artificial Neural Networks in Studies of Steels and Alloys), Vol. 11(3), pp. 388, (2021). doi.org/10.3390/met11030388
Dang, T.L., Cao, T., Hoshino, Y.: Classification of Metal Objects Using Deep Neural Networks in Waste Processing Line. Int. J. Innovative Comput. Inform. Control. 15(5), 1901–1912 (2019)
Kociołek, M., Kozłowski, M., Cardone, A.A.: Convolutional Neural Networks-Based Approach for Texture Directionality Detection. Sensors. 22, 562 (2022). https://doi.org/10.3390/s22020562
Eftekhar Hossain, M.F., Hossain, M.A., Rahaman, A., Color: and Texture Based Approach for the Detection and Classification of Plant Leaf Disease Using KNN Classifier, 2019 International Conference on Electrical, Computer and Communication Engineering (ECCE), 7–9 February, (2019)
Turgut, Ozseven: Infant cry classification by using different deep neural network models and hand-crafted features. Biomed. Signal Process. Control. 83, 1746–8094 (2023). https://doi.org/10.1016/j.bspc.2023.104648
Małgorzata Szafraniec, Z., Omiotek: Danuta Barnat-Hunek, Water absorption prediction of nanopolymer hydrophobized concrete surface using texture analysis and machine learning algorithms. Constr. Build. Mater. 375 (2023). 130969, ISSN 0950 – 0618 https://doi.org/10.1016/j.conbuildmat.2023.130969
Pan, Z., Liu, Y., Li, Z., Xun, Q., Wu, Y.: Roughness detection method based on image multi-features. Proceedings of the Institution of Mechanical Engineers, Part E: Journal of Process Mechanical Engineering. doi: (2023). 10.1177/09544089231154959
Dillam Jossue Díaz-Romero, Van den Eynde, S., Sterkens, W., Engelen, B., Zaplana, I., Dewulf, W., Goedemé, T., Peeters, J.: Simultaneous mass estimation and class classification of scrap metals using deep learning, Resources, Conservation and Recycling,Vol. 181,2022, 106272, https://doi.org/10.1016/j.resconrec.2022.106272

No competing interests reported.

Classification of Metal and Metal Oxide Nanoparticles Using Machine Learning and Deep Learning Models

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and Methods

3. Proposed Method

4. Experimental Results and Discussion

4.1 Machine learning:

4.2 Deep learning:

5. Conclusion

Declarations

Acknowledgments

References

Additional Declarations

Status:

Version 1