Ink Classification Using Convolutional Neural Network

Ink Classification Using Convolutional Neural

Network

Binu Melit Devassy

_{and Sony George}

1 _{Department of Computer Science, Norwegian University of Science and Technology} [email protected], [email protected]

Abstract

Ink classification plays a vital role in document forgery detection and analysis. Hyperspectral imaging (HSI) is an emerging tool for non-invasive analysis of materials in various fields including forensic analysis. HSI records hundreds of narrowband images in electromagnetic spectrum and this will produce larger data sizes compared to conventional imaging techniques. In order to handle this large information, it is necessary to develop some automated and more computationally efficient methods for data analysis. This paper presents the deep learning approach for hyperspectral ink classification by using one-dimensional (1D) Convolutional Neural Network (CNN). In addition to that, we have compared the CNN results against two standard hyperspectral classification methods; Spectral Angle Mapper (SAM) and Spectral Information Divergence (SID). The results from classification algorithms revealed the effectiveness of CNN for ink classification with respect to the two conventional methods used.

1

Introduction

Identification of several ink types present in a questioned document can be a significant evidence in document forgery detection; this might be inferred as an evidence related to the manipulations taken place. Most of the existing tools to identify forged or altered document are sophisticated and time consuming. Usually document forensic laboratory uses microscopes, various imaging equipment that can capture images under ultraviolet (UV) and infrared (IR) light sources and resources to accomplish chemical analysis[1]. Visual inspection to identify some information regarding fictitious work is possible however, this is subjective and also influenced by several external parameters. To achieve a trustable evidence, the analyst requires a reliable technology. The methods used in document analysis can be classified as destructive and non-destructive[2], document analysts prefer to use non-destructive methods rather than destructive to save the evidence even after investigation[3]. The major non-destructive techniques used in document and ink analysis are, Fourier transform infrared (FTIR)[4], Raman spectroscopy[1], video spectral comparator (VSC)[5], multi-spectral imaging[6] and

hyperspectral imaging (HSI)[7]. However, chromatography and electrophoresis[8] have been considered as the potential destructive methods for ink analysis.

HSI combines conventional imaging and spectroscopy to obtain the spectral information over the entire spatial areas of interest[9]. HSI images are comprised of hundreds of narrow band images across the spatial region in the target. Due to this abundant information and non-destructive nature of HSI images, HSI become an increasingly adopted technique in many fields including food quality[10], medical imaging[11], material science[12], cultural heritage[13] and forensics[14]. HSI images are represented in three-dimensional(3D) format in comparison with two-dimensional(2D) image and can be called as “hypercube”[15]. The spatial dimensions are along vertical and horizontal directions of each spectral slice in the cube and the third dimension is along supported spectral rangeas illustrated in Fig. 1. Each pixel of HSI data can be expressed as a spectrum in that particular position (Fig.1) and this spectrum can be used as a fingerprint for characterizing each pixel.

Figure 1 Hyperspectral data structure

As a rapid non-destructive and non-contact method, HSI offers significant potential for the detection, visualization, identification and age estimation of forensic traces[14]. The review report made by G.J.Edelman et al.[14] described many applications of HSI in forensic analysis including estimating age of bruises[16], age estimation of blood stains at the crime scene[17] and fingerprint analysis[18]. The study done by Zhihong Pan et al.[19] demonstrated an algorithm using HSI can be used to recognize faces over time in the presence of changes in facial pose and expression. In document analysis; Aythami Morales et al. proposed a method for ink analysis in handwritten document using HSI[20], also G.Reed et al. explored the power of HSI for discriminating gel pen inks[21] and this report[22] shows the potential of HSI to distinguish documents having highly unbalanced ink proportions. Another important work by Butt et al. [23], employed HSI imaging for automatic signature extraction. Wang et al.[24] developed a method to identify tampered handwriting using HSI and recently Muhammad et al.[25] explored the potential of HSI for analyzing gel pen inks. A much closer previous work is presented in this report[26], which compared few important classification methods on hyperspectral ink data. The previous work using convolutional neural networks (CNN) was by MJ Khan et al.[27] used a 2D-CNN for automated forgery detection. Their method used the data having 33 bands from 5 blue and 5 black pens, however this experiment used data having 186 bands from 29 blue pens, hence it is not fair to have a direct comparison with this 2D-CNN method against the proposed method in this work.

Classification of hyperspectral image is a difficult task because of high dimensionality of the spectral data, limited number of training samples and the presence of redundant features[28]. The review report made by P. Ghamisi et al.[28] described the challenges and various classification methods in HSI classification. The HSI classification methods can be classify into four major groups and are numerical methods, probabilistic based methods, kernel methods and deep learning methods. The numerical methods are Euclidean distance[29], spectral angle mapper (SAM)[30], spectral correlation mapper (SCM)[30] etc., probabilistic methods are spectral information divergence (SID)[31], multinomial logistic regression[32] etc.. Support-vector machines (SVM)[33] are the major candidate in kernel based methods and Yushi Chen et al.[34] proposed a joint spectral-spatial deep neural network for hyperspectral classification.

In this work, we have introduced and evaluated one-dimensional (1D) neural network into hyperspectral ink classification and compared the result against standard hyperspectral classifiers such as SAM and SID. We selected 1D convolutional neural network (CNN) because of its advantages over two-dimensional (2D) CNN, which will be discussed later in this article (Section 2.5). The remaining part of this report is organized as three parts; the first part will explain the sample preparation, HSI data acquisition, 1D CNN and evaluation methods. In the last section we will discuss results, followed by conclusion and possible future works.

2

Materials and Methods

2.1

Sample Preparation

Sample texts were prepared from twenty nine blue pens, these pens are from different brands and ink types such as gel ink, ballpoint and liquid ink. Table 1 shows the details of pens and brands used. Most of the pens used in the present study are available internationally; this will help to reproduce the dataset for ensuring the repeatability of the experiment. The samples dried for twenty four hours before HSI acquisition. The pens were marked with labels from one to twenty nine, for example Pen 1, Pen 2 etc. same labels are used throughout this paper. The popular saying “All roads lead to Rome” written by same person is used as sample text for all pens.

Id Brand Model Type Id Brand Model Type 1 Flair Sunny Ball 16 Pentel Energel Gel 2 Pilot Frixion Gel 17 Pilot

Hi-Techpoint V5 Rt

Ink

3 Epoca Svart Ball 18 BIC Unknown Ball

4 Uni Jetstream Gel

19 Unknown

Marketing Pen Ball 5 Papermate Ink Joy Ball

20 Unknown

Marketing Pen Ball 6 Pentel Energel Gel 21 Frixion Ball Slim

038 Gel

7 Prodir Z+F

Zoller Ball 22 BIC

Cristal Medium Blue

Ball

8 Pilot Hi

Tecpoint Ink 23 Pilot Frixion Gel 9 Schneider Xpress Ink 24 Pilot Rexgrip Ball 10 Cello Papersoft Ball 25 Uni Power

Tank Ball 11 Schneider Tops 505

M Ball 26 Ntnu Ball

12 Pilot G-Knock Gel 27 Unknown Unknown Ball 13 Unknown Marketing

Pen Ball 28 Biltema Unknown Gel 14 Beifa Liquidly Ink 29 Paper

Mate Replay Ball

15 Flair Unknown Ball

Table 1 Details of pens used in the present study

2.2

Hyperspectral Acquisition

The HySpex VNIR-1800[35] push broom camera from Norsk Elektro Optikk AS is used for HSI acquisition. The camera was placed perpendicular to the translator stage where the sample paper with hand written text was placed. Two halogen light sources were placed at an angle of 45 degrees from the camera as shown in Fig. 2. VNIR 1800 can capture 186 spectral bands across wavelength starting from 400 to 1000 nm and has a spectral sampling of 3.18 nm with a spatial resolution of 1800 pixels across the field of view of 10 cm. A Contrast Multi-Step Target[36] was used as reference, which was present in the scene. The camera software HySpex RAD converts the raw images into the sensor absolute radiance values and performs radiometric calibrations.

Figure 2 Hyperspectral acquisition setup

It is better have an outlook of the data set used in this experiment. The Fig.3 illustrates the mean spectrums of twenty-nine pens used in this experiment. From these spectrums, it is clear that many pen inks having nearly similar spectral signatures. Hence, it would be obvious that, identification of each spectra will be a challenging task without a powerful classifier.

2.3

Spectral Angle Mapper (SAM)

The SAM is a well-known spectral classification tool for measuring the shape similarity between the reference and test spectra. The SAM algorithm calculates the angle between reference and test spectrum by treating them as vectors in a space with dimensionality equal to the number of bands (nb)[29].

Figure 4 A two band image, alpha (α) represents the angle between the vectors corresponding to test and reference spectrum.

The angle alpha (α) defines the similarity between test and reference spectrum, alpha can be calculated using the below equation (Equation 1), where r and t are the reference and target spectrums.

𝛼 = cos−1₍ ∑𝑛𝑏𝑖=1𝑡𝑖𝑟𝑖 (∑𝑛𝑏_𝑖=1𝑡_𝑖2) 1 2 ⁄ (∑𝑛𝑏_𝑖=1𝑟_𝑖2) 1 2 ⁄) (1)

2.4

Spectral Information Divergence (SID)

Unlike SAM which compares the geometric features, SID consider each pixel in the spectrum as a random variable and finds the similarity between reference and target spectra by measuring the discrepancy between probabilistic behaviors of two spectra[31]. The spectrums will be normalized to a range of [0,1], by defining 𝑝 =_∑𝑎𝑗

𝑎𝑖 𝐿 𝑖=1

. Here ‘a’ is the given hyperspectral pixel vector a = (a1,---,aL)T, then define a normalized vector p= {𝑝}𝑖=1𝐿 . Similarly we can define reference(x) and test(y)

spectrum as vectors ‘r’ and ‘t’ , then SID can be define as below 𝑆𝐼𝐷(𝑥, 𝑦) = 𝐷(𝑥 ∥ 𝑦 ) + 𝐷(𝑦 ∥ 𝑥 )

Where

and

2.5

Convolutional Neural Networks (CNN)

Convolutional Neural Networks (CNNs) were proven very effective in image classification and reported promising results for 1D signals such as audio[37], Electrocardiographic (ECG) signals[38]

and other 1D time signals[39][40]. Hence, it is sensible to use 1D-CNN for spectral classifications, because spectral signatures are considered as 1D signals with discrete values along supported wavelength. In addition, 1D-CNNs have certain advantage over 2D-CNNs, they are mentioned below[41]

 Computational complexity of CNNs are significantly lower than that of 2D-CNNs, 1D-CNNs requires simple array operations.

 1D-CNNs requires relatively shallow architectures, however 2D-CNNs usually require deeper architectures.

 Because of the two reasons mentioned above, 1D-CNNs requires regular hardware for training, but 2D- CNNs quite powerful hardware for training.

 Due to the above-mentioned facts 1D-CNNs are well suited for real-time and low-cost applications.

The architecture used in this experiment is illustrated in Fig.5. The hidden layer consists of n-number of 1D-CNN and ReLU pairs; the ‘n’ will be decided after training with possible values which gives maximum accuracy. A dropout layer with rate ‘0.5’ is present after hidden layers and then a max pooling layer with a pool size of ‘2’. The fully connected layers will follow the max-pooling layer, then the softmax layer, which gives the output classifications. The network used Adam[42] optimization and the categorical cross-entropy as loss function, which is well known for learning a multi-class classification problem.

Figure 5 CNN architecture used in this study

2.6

Data Processing and Training

Acquired HSI data were normalized using the standard reference target, which was present in the scene. Out of nine-hundred individual spectrums extracted for each pen, 80% of the spectrums were used for training and the remaining for the evaluation. For SAM and SID, 80% training data were used to estimate the reference spectrum by taking the mean spectra and 20% data for evaluation. The entire experiment framework was implemented in Python and Keras[43]. For parameter tuning, SHERPA[44] was used, a Python library for hyper parameter tuning of machine learning models. All the parameters of CNN models are initialized with random Gaussian distributions and used Bayesian optimization for hyper parameter tuning[45]. The parameters included in tuning are; number of filters and kernel size

for convolution layer, batch size for training, learning rate, number of hidden layers and number of epochs.

2.7

Accuracy Calculation

After parameter tuning the accuracy and other performance indexes were calculated from confusion matrix[46] as follows

 Accuracy: The ratio between truly predicted outcomes (TP+TN) and the sum of all predictions (TP+TN+FP+FN).

 Precision: The ratio between all true positives (TP) and all positive predictions (TP+FP).  Recall: The ratio between true positives (TP) divided by all actual positives (TP+FN).  F1-score: The weighted average of precision and recall.

The terms TP, TN, FN and FP used in the above definitions are explained in the Fig. 6

Figure 6 Actual vs predicted outcomes

3

Results and Discussion

3.1

Results

With parameter tuning, we have identified the optimal parameters for this task and are presented in the Table 2. Obtained accuracy of 91% on test data, which was very close to the training accuracy of 92.9%. Because of the 1D kernel with size ‘3’ and low filter count (16) the training was pretty fast and finished in 307 seconds and test finished in 0.56 seconds on Intel Core i7 8650U CPU with 16 GB RAM without a dedicated GPU.

Parameter Value Range Used Optimal Value

Kernal Size [3, 5,7,9,11,13,15] 3

Filter Count [8, 16, 32, 64, 128, 256] 16

Number of hidden layers In range between 2 to 10 4

Epochs In range between 5 to 50 35

Batch size [32, 64, 128, 256] 32

Learning rate In range between 0.001 to 0.1 0.00464

Table 2 Parameter used for tuning and the optimal values obtained

The more detailed CNN results for each ink are illustrated in Table 3, which is created from the confusion matrix generated from the test results. This table also includes precision obtained from SAM and SID and these values placed next to that of CNN for each pen to make comparison easier. As we are focusing to the CNN results, excluded other performance indexes for SAM and SID. The overall

test accuracies are given in the last row, from overall accuracies it is clear that CNN method outperformed other methods. Figures 7 to 9 illustrated the confusion matrix obtained for each classification method; it will show the exact number of wrong and correct predictions generated by each method.

Pen IDs Precision SID Precision SAM Precision CNN Recall CNN F1-score CNN Support CNN 1 0.66 0.61 0.94 0.95 0.94 180 2 0.26 0.32 0.76 0.79 0.77 180 3 0.15 0.21 1 0.99 0.99 180 4 0.55 0.71 1 1 1 180 5 0.32 0.36 1 0.6 0.75 180 6 0.49 0.45 0.7 0.76 0.73 180 7 0.72 0.73 0.99 1 1 180 8 0.59 0.38 1 0.97 0.99 180 9 0.46 0.84 1 1 1 180 10 0.32 0.36 0.58 0.39 0.47 180 11 0.62 0.72 0.98 0.98 0.98 180 12 0.53 0.55 0.97 0.98 0.98 180 13 0.47 0.48 0.99 1 0.99 180 14 0.58 0.67 1 1 1 180 15 0.29 0.33 0.56 0.97 0.71 180 16 0.5 0.44 0.73 0.69 0.71 180 17 0.41 0.62 0.97 0.97 0.97 180 18 0.29 0.32 1 0.98 0.99 180 19 0.3 0.33 0.98 0.89 0.93 180 20 0.58 0.53 0.94 0.99 0.96 180 21 0.47 0.49 0.73 0.91 0.81 180 22 0.56 0.6 0.98 0.99 0.99 180 23 0.47 0.45 0.83 0.58 0.68 180 24 0.42 0.56 1 0.99 1 180 25 0.56 0.75 0.99 0.99 0.99 180 26 0.23 0.26 0.94 0.97 0.95 180 27 0.31 0.37 0.95 0.98 0.96 180 28 1 1 1 1 1 180 29 0.59 0.47 1 0.96 0.98 180 SAM SID CNN Accuracy 0.49 0.43 0.91

Figure 7 Confusion matrix obtained from 1D-CNN results

Figure 9 Confusion matrix obtained from SID results

3.2

Discussion

As we seen in the result section, SAM and SID has very low accuracy values than that of CNN method and it can be explained based on the way how they works. The SAM reveals the similarity between reference and test spectra by calculating the spectral angle between n dimensional vectors in a space; where n represents the number of bands[47]. This method is relatively insensitive to slight variations in spectrum, hence SAM might produce a zero spectral angle even when the two vectors are not identical[48]. However, this strength introduced a weakness into SAM; it might introduce wrong predictions between nearly similar spectrums. The present study has few pens with nearly close spectrums, for example consider the pen 6 and 16 they possess nearly similar spectral shape, which is illustrated in Fig.10. In addition, the figure shows the SAM angle and it was pretty close to zero hence SAM misinterpreted 6 as 16. The SAM confusion matrix in Fig.8 will give the exact count that how many times SAM interpreted pen 6 as 16, it was 80 out of total 180 samples.

However, SID is designed to extract the discrepancy between the probability distributions of the spectrums[49]. Even though SID designed for solving weaknesses of SAM, it also depends on shape similarity between the spectral signatures. Hence, SID also produced wrong decisions for nearly similar spectral inputs and which can be proven from the Fig.10, for pen 6 and 16 the SID score was ‘0’ means for SID both spectrums seems same. The SID confusion matrix in Fig.9 will give a clearer picture about the depth of error in the classification; it was 101 wrong predictions out of 180 input samples. However, in CNNs each filters will extract some information related to the spectrum under process and finally combines all these information to make a final design; this chunk of information will help CNNs to learn and later execute classification. SAM and SID are well-known and proven their classification power in hyperspectral remote sensing applications[50][51], but obtained poor performance score for ink classification task.

To verify our justification for the low accuracy obtained by SAM and SID, a subset of the original dataset containing 10 pens were created by excluding pens having similar spectrums. Training and testing were performed on this dataset for 1D-CNN, SAM and SID. Resulted accuracies are presented in Table 4 and the spectrum used for this purpose is illustrated in Fig.11. From Table 4, it is clear that the accuracies of SAM and SID improved further, because of less confusing spectral pairs. Also, the proposed method improved its accuracy to 98% compared to 91% with full dataset.

SAM SID CNN

Accuracy 0.74 0.7 0.98

Figure 11 Manually selected pen spectrums

4

Conclusion

The 1D-CNN is introduced into hyperspectral ink data analysis. This experiment used hyperspectral data from 29 blue pens of different types. The results were compared against standard classification methods such as SAM and SID, the proposed 1D-CNN method outperformed SAM and SID in ink classification. As the proposed method uses 1D-CNN which helped to obtain faster training and real time test results, which will support the deployment of these kind of solutions into mobile devices. This experiment used only 29 pens, however testing with more pens will be considered as a potential future work. The proposed method used the basic 1D-CNN to classify pens, however this can be extended using more complex and deeper neural network that are already proven for other classification tasks. ACKNOWLEDGEMENT

The authors would like to offer their thanks to Norsk Elektro Optikk, Norway for their support in conducting this research.

References

[1] F. de Souza Lins Borba, R. Saldanha Honorato, and A. de Juan, “Use of Raman spectroscopy and chemometrics to distinguish blue ballpoint pen inks,” Forensic Sci. Int., vol. 249, pp. 73– 82, 2015.

[2] C. S. Silva, F. de S. L. Borba, M. F. Pimentel, M. J. C. Pontes, R. S. Honorato, and C. Pasquini, “Classification of blue pen ink using infrared spectroscopy and linear discriminant analysis,”

Microchem. J., vol. 109, pp. 122–127, 2013.

[3] A. Braz, M. López-López, and C. García-Ruiz, “Raman spectroscopy for forensic analysis of inks in questioned documents,” Forensic Science International, vol. 232, no. 1–3. pp. 206–212, 2013.

synchrotron reflectance infrared spectromicroscopy as a rapid, direct, nondestructive method for the study of inks on paper,” Appl. Spectrosc., vol. 56, no. 6, pp. 800–803, 2002.

[5] R. Kaur, K. Saini, and N. C. Sood, “Application of Video Spectral Comparator (absorption spectra) for establishing the chronological order of intersecting printed strokes and writing pen strokes,” Sci. Justice, vol. 53, no. 2, pp. 212–219, 2013.

[6] J. Havermans, H. A. Aziz, and H. Scholten, “Non destructive detection of iron gall inks by means of multispectral imaging. Part 1: Development of the detection system,” Restaurator, vol. 24, no. 1, pp. 55–60, 2003.

[7] S. George and J. Y. Hardeberg, “Ink classification and visualisation of historical manuscripts: Application of hyperspectral imaging,” in Proceedings of the International Conference on Document Analysis and Recognition, ICDAR, 2015, vol. 2015–Novem, pp. 1131–1135. [8] J. A. Zlotnick and F. P. Smith, “Chromatographic and electrophoretic approaches in ink

analysis,” Journal of Chromatography B: Biomedical Sciences and Applications, vol. 733, no. 1–2. pp. 265–272, 1999.

[9] A. A. Gowen, C. P. O’Donnell, P. J. Cullen, G. Downey, and J. M. Frias, “Hyperspectral imaging - an emerging process analytical tool for food quality and safety control,” Trends Food Sci. Technol., vol. 18, no. 12, pp. 590–598, 2007.

[10] G. ElMasry, M. Kamruzzaman, D. W. Sun, and P. Allen, “Principles and Applications of Hyperspectral Imaging in Quality Evaluation of Agro-Food Products: A Review,” Critical Reviews in Food Science and Nutrition, vol. 52, no. 11. pp. 999–1023, 2012.

[11] G. Lu and B. Fei, “Medical hyperspectral imaging: a review,” J. Biomed. Opt., vol. 19, no. 1, p. 10901, 2014.

[12] P. Tatzer, M. Wolf, and T. Panner, “Industrial application for inline material sorting using hyperspectral imaging in the NIR range,” Real-Time Imaging, vol. 11, no. 2, pp. 99–107, 2005. [13] C. Fischer and I. Kakoulli, “Multispectral and hyperspectral imaging technologies in conservation: current research and potential applications,” Stud. Conserv., vol. 51, no. sup1, pp. 3–16, 2006.

[14] G. J. Edelman, E. Gaston, T. G. van Leeuwen, P. J. Cullen, and M. C. G. Aalders, “Hyperspectral imaging for non-contact analysis of forensic traces,” Forensic Science International, vol. 223, no. 1–3. pp. 28–39, 2012.

[15] R. Lu and Y.-R. Chen, “Hyperspectral imaging for safety inspection of food and agricultural products,” in Proc.SPIE, 1999, vol. 3544.

[16] G. Payne, N. Langlois, C. Lennard, and C. Roux, “Applying visible hyperspectral (chemical) imaging to estimate the age of bruises,” Med. Sci. Law, vol. 47, no. 3, pp. 225–232, 2007. [17] G. Edelman, T. G. van Leeuwen, and M. C. G. Aalders, “Hyperspectral imaging for the age

estimation of blood stains at the crime scene,” Forensic Sci. Int., vol. 223, no. 1–3, pp. 72–77, 2012.

[18] E. Bartick, R. Schwartz, R. Bhargava, and M. Schaeberle, “Spectrochemical Analysis and Hyperspectral Imaging of Latent Fingerprints,” Proc. 16th Meet. Int. Assoc. Forensic Sci. Montpellier Fr., pp. 61–64, 2002.

[19] Z. Pan, G. Healey, M. Prasad, and B. Tromberg, “Face Recognition in Hyperspectral Images,”

IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 12, pp. 1552–1560, 2003.

[20] A. Morales, M. A. Ferrer, M. Diaz-Cabrera, C. Carmona, and G. L. Thomas, “The use of hyperspectral analysis for ink identification in handwritten documents,” in Proceedings - International Carnahan Conference on Security Technology, 2014, vol. 2014–Octob, no. October.

[21] G. Reed, K. Savage, D. Edwards, and N. Nic Daeid, “Hyperspectral imaging of gel pen inks: An emerging tool in document analysis,” Sci. Justice, vol. 54, no. 1, pp. 71–80, 2014.

[22] Z. Luo, F. Shafait, and A. Mian, “Localized forgery detection in hyperspectral document images,” in Proceedings of the International Conference on Document Analysis and

Recognition, ICDAR, 2015, vol. 2015–Novem, pp. 496–500.

[23] U. M. Butt, S. Ahmad, F. Shafait, C. Nansen, A. S. Mian, and M. I. Malik, “Automatic signature segmentation using hyper-spectral imaging,” in Proceedings of International Conference on Frontiers in Handwriting Recognition, ICFHR, 2017, pp. 19–24.

[24] W. Wang, L. Zhang, D. Wei, Y. Zhao, and J. Wang, “The principle and application of hyperspectral imaging technology in detection of handwriting,” in 2017 9th International Conference on Advanced Infocomm Technology, ICAIT 2017, 2018, pp. 345–349.

[25] M. N. M. Asri, N. A. M. Noor, W. N. S. M. Desa, and D. Ismail, “Forensic Analysis of Gel Pen Inks using Hyperspectral Imaging coupled with Chemometric Procedures,” Arab J. Forensic Sci. Forensic Med., vol. 1, no. 9, pp. 1174–1184, 2019.

[26] B. M. Devassy, S. George, and J. Y. Hardeberg, “Comparison of Ink Classification Capabilities of Classic Hyperspectral Similarity Features,” in 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), 2019, pp. 25–30.

[27] M. J. Khan, “Deep learning for automated forgery detection in hyperspectral document images,” J. Electron. Imaging, vol. 27, no. 5, p. 1, 2018.

[28] P. Ghamisi, J. Plaza, Y. Chen, J. Li, and A. J. Plaza, “Advanced Spectral Classifiers for Hyperspectral Images: A review,” IEEE Geoscience and Remote Sensing Magazine, vol. 5, no. 1. pp. 8–32, 2017.

[29] N. Keshava, “Distance metrics and band selection in hyperspectral processing with applications to material identification and spectral libraries,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 7, pp. 1552–1565, 2004.

[30] O. . De Carvalho and P. R. Meneses, “Spectral correlation mapper (SCM): an improvement on the spectral angle mapper (SAM). In: Summaries of the 9th JPL airborne earth science workshop,” Summ. 9th JPL Airborne Earth Sci. Work., no. JPL Publication 00-18, 9 , 2000. [31] C. I. Chang, “Spectral information divergence for hyperspectral image analysis,” in

International Geoscience and Remote Sensing Symposium (IGARSS), 1999, vol. 1, pp. 509– 511.

[32] J. Li, J. M. Bioucas-Dias, and A. Plaza, “Semisupervised hyperspectral image segmentation using multinomial logistic regression with active learning,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 11, pp. 4085–4098, 2010.

[33] D. Tuia and G. Camps-Valls, “Semisupervised remote sensing image classification with cluster kernels,” IEEE Geosci. Remote Sens. Lett., vol. 6, no. 2, pp. 224–228, 2009.

[34] Y. Chen, Z. Lin, X. Zhao, G. Wang, and Y. Gu, “Deep learning-based classification of hyperspectral data,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 7, no. 6, pp. 2094– 2107, 2014.

[35] “HySpex VNIR 1800.” [Online]. Available: https://www.hyspex.no. [Accessed: 31-Aug-2019]. [36] “Contrast Multi-Step Target.” [Online]. Available: https://www.labspherestore.com/.

[37] S. Hershey et al., “CNN architectures for large-scale audio classification,” in ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings, 2017, pp. 131–135.

[38] R. Donida Labati, E. Muñoz, V. Piuri, R. Sassi, and F. Scotti, “Deep-ECG: Convolutional Neural Networks for ECG biometric recognition,” Pattern Recognition Letters, 2018.

[39] O. Abdeljaber, O. Avci, S. Kiranyaz, M. Gabbouj, and D. J. Inman, “Real-time vibration-based structural damage detection using one-dimensional convolutional neural networks,” J. Sound Vib., vol. 388, pp. 154–170, 2017.

[40] G. Manganaro and J. P. De Gyvez, “One-dimensional discrete-time CNN with multiplexed template-hardware,” IEEE Trans. Circuits Syst. I Fundam. Theory Appl., vol. 47, no. 5, pp. 764–769, 2000.

[41] S. Kiranyaz, O. Avci, O. Abdeljaber, T. Ince, M. Gabbouj, and D. J. Inman, “1D Convolutional Neural Networks and Applications: A Survey,” May 2019.

[42] D. P. Kingma and J. L. Ba, “Adam: A method for stochastic gradient descent,” ICLR Int. Conf. Learn. Represent., 2015.

[43] “Keras.” [Online]. Available: https://keras.io/. [Accessed: 02-Sep-2019].

[44] “Sherpa.” [Online]. Available: https://parameter-sherpa.readthedocs.io/en/latest/index.html. [Accessed: 02-Sep-2019].

[45] B. J. Snoek, H. Larochelle, R. P. Adams, and J. Snoek, “PRACTICAL BAYESIAN OPTIMIZATION OF MACHINE LEARNING,” Arxiv, pp. 1–12, 2001.

[46] T. Fawcett, “An introduction to ROC analysis,” Pattern Recognit. Lett., vol. 27, no. 8, pp. 861– 874, 2006.

[47] F. Kruse, “Techniques developed for geologic analysis of hyperspectral data applied to nearshore hyperspectral ocean data,” … Conf. Remote …, no. March 1997, pp. 17–19, 1997. [48] M. N. Kumar et al., “A new hybrid spectral similarity measure for discrimination among vigna

species,” Int. J. Remote Sens., vol. 32, no. 14, pp. 4041–4053, 2011.

[49] C. I. Chang, “An information-theoretic approach to spectral variability, similarity, and discrimination for hyperspectral image analysis,” IEEE Trans. Inf. Theory, vol. 46, no. 5, pp. 1927–1932, 2000.

[50] E. L. Hunter and C. H. Power, “An assessment of two classification methods for mapping Thames Estuary intertidal habitats using CASI data,” Int. J. Remote Sens., vol. 23, no. 15, pp. 2989–3008, 2002.

[51] F. van der Meer, “The effectiveness of spectral similarity measures for the analysis of hyperspectral imagery,” Int. J. Appl. Earth Obs. Geoinf., vol. 8, no. 1, pp. 3–17, 2006.