Source printer identification using convolutional neural network and transfer learning approach

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 13, No. 1, March 2024, pp. 948~960
ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i1.pp948-960  948
Journal homepage: http://ijai.iaescore.com
Source printer identification using convolutional neural
network and transfer learning approach
Naglaa F. El Abady1
, Hala H. Zayed1,2
, and Mohamed Taha1
1
Department of Computer Science, Faculty of Computers and Artificial Intelligence, Benha University, Benha, Egypt
2
School of Information Technology and Computer Science (ITCS), Nile University, Giza, Egypt
Article Info ABSTRACT
Article history:
Received Jan 20, 2023
Revised May 27, 2023
Accepted Jun 3, 2023
In recent years, Source printer identification has become increasingly
important for detecting forged documents. A printer's distinguishing feature
is its fingerprints. Each printer has a unique collection of fingerprints on every
printed page. A model for identifying the source printer and classifying the
questioned document into one of the printer classes is provided by source
printer identification. A paper proposes a new approach that trains three
different approaches on the dataset to choose the more accurate model for
determining the printer's source. In the first, some pre-trained models are used
as feature extractors, and support vector machine (SVM) is used to classify
the generated features. In the second, we construct a two-dimensional
convolutional neural network (2D-CNN) to address the source printer
identification (SPI) problem. Instead of SoftMax, 2D-CNN is employed for
feature extractors and SVM as a classifier. This approach obtains 93.75%
98.5% accuracy for 2D-CNN-SVM in the experiments. The SVM classifier
enhanced the 2D-CNN accuracy by roughly 5% over the initial configuration.
Finally, we adjusted 13 already-pre-trained CNN architectures using the
dataset. Among the 13 pre-trained CNN models, DarkNet-19 has the greatest
accuracy of 99.2 %. On the same dataset, the suggested approaches achieve
well in terms of classification accuracy than the other recently released
algorithms.
Keywords:
Convolution neural network
Document forgery
Source printer identification
Transfer learning
Two-dimensional convolution
neural network
This is an open access article under the CC BY-SA license
Corresponding Author:
Naglaa F. El Abady
Department of Computer Science, Faculty of Computers and Artificial Intelligence, Benha University
Benha Mansoura Road, next to Holding Company for Water Supply and Sanitation Benha
Qalyubia Governorate, Egypt
Email: naglaa.fathy@fci.bu.edu.eg
1. INTRODUCTION
The process of examining and analyzing digital evidence to determine a crime's occurrence is known
as digital forensics. To locate, collect, and analyze digital evidence, digital forensics requires specialized tools
and processes [1]. The use of digital documents has exploded in the recent decade. These digital papers could
include images of official contracts, bills, checks, and so on. The goal of these digital papers is to eliminate the
need for paper. Compared to a hard copy, maintaining a digital document is also simpler, cheaper, and more
effective, but safety is a problem.
Personal computers, scanners, and printers can produce forged documents such as certificates,
agreements, identity cards, and lottery tickets. Modern printers have such high resolution that it is difficult for
normal persons to differentiate forged documents from real ones. Traditional approaches use chemical
techniques to detect forgeries in printed documents [2], [3]. These procedures need laboratory tools and a
specialist to evaluate the samples. Additionally, these methods take a long time and risk damaging the printed

Int J Artif Intell ISSN: 2252-8938 
Source printer identification using convolutional neural network and transfer … (Naglaa F. El Abady)
949
paper. Digital techniques, in contrast, use a reference scanner to turn printed papers into their digital equivalent.
Using digital approaches for source printer identification, it is possible to distinguish between documents
printed on various printers. Since all the analysis is done digitally, it is quicker and more automatic.
Extrinsic (active) and intrinsic (passive) are the two main digital methods for identifying source
printers. Prior research typically used the following techniques to extract statistical features from printed
documents: Discrete wavelet transform (DWT), local binary pattern (LBP), key printer noise features (KPNF),
gray-level co-occurrence matrix (GLCM), speeded up robust features (SURF), orientated FAST rotated and
BRIEF (ORB), Histogram of gradient (HOG) and spatial filters [4]–[10], [11]–[16]. Subsequently, the forensic
classification system adopts support vector machine (SVM), random forest (RF), and ensemble techniques.
However, the tasks involving feature extraction, feature selection, and classification, as indicated in the
preceding statement, demand much professional human participation. Additionally, to obtain results that may
be generalized, the entire procedure must be repeated several times using a random selection of training and
testing samples. Moreover, machine learning methods have fundamentally altered the research area for both
academia and industry. A branch of artificial intelligence is machine learning (AI) [7]. The general goal of
machine learning is to identify the structure of data and fit it into models that are both useful and
understandable.
Deep learning needs a large number of datasets to give more accuracy. In this research, we solve the
problem of the large dataset by using transfer learning. The transfer learning convolutional neural network
(CNN) was used to save time for creating a model from scratch, training, and small datasets. Transfer learning
CNN is divided into two types: the first used a pre-trained model for feature extraction, and the result was
classified using machine learning classification techniques. The second used a pre-trained model but changed
the last layers. We also used 2D-CNN models with SoftMax or SVM classifiers to determine the printer's
source. The significant contributions of this work include the following,
− Up to our experience, this is the first approach using CNN for source printer identification (SPI) of full
printed text documents without dividing the document into letters, words, or patches.
− Multiple pre-trained CNN models are modified to increase the efficiency of detecting forgery in
documents based on the SVM classifier instead of SoftMax.
− We successfully developed a pre-processing stage that included histogram equalisation and gamma
correction, which significantly improved the model's performance and accuracy.
− The proposed approach performs better in classification accuracy than the other recently published
algorithms on the same dataset.
The following sections make up the entire paper. While section 3 focuses on the background,
section 2 provides a short description of the related work for classifying the source printer of a printed text
document. Section 4 contains a description of the specifics of our proposed approach. The efficiency of the
proposed approach is investigated using a detailed series of experiments. The proposed approach's description
and outcomes have been explored in section 5. Finally, conclusions from this effort are presented in section 6.
2. RELATED WORK
Document manipulation can be detected using a variety of approaches. The bulk of these approaches
detects the printer's source to identify the printer types employed in the printing process. The issue of source
printer (SP) classification has received a lot of attention in the last decade [17]. This section will go over the
most popular ways for authenticating a document and ensuring that it was printed by a legal printer.
Mikkilineni et al. [4], proposed a printer identification procedure based on an SVM classifier. They investigated
the impact of font size, font type, paper type, and printer age. When the variables font size, paper type, and
printer age are constant, their printer identification technique works. Jung et al. [5], introduced a novel colour
laser printer forensic algorithm. It uses an SVM classifier and noisy texture analysis. To estimate invisible
noises, the Wiener filter and the 2D DWT filter are used. The GLCM is then applied to the noise texture. The
machine classifier is trained and validated using data from 384 statistical features. For brand, toner, and model
recognition, the presented technique achieves 99.3%, 97.4%, and 88.7% accuracy, respectively. The authors
in [6], described a method for identifying document forgeries based on distortion mutation of geometric
parameters (DMGP) with translation and rotation distortion parameters. This technique can be used to examine
papers in both English and Chinese. It can analyze papers based on their individual characters. It is resistant to
JPEG compression and performs well with low-resolution documents. The GLCM and DWT were used to
extract texture features from the Chinese printed source to evaluate the effects of different output devices [7].
The feature selection techniques are utilized to select the most appropriate feature subset and an SVM to
determine the source model of the documents. The overall experimental results achieve an identification rate
of 98.64%, 1.27% higher than the GLCM method that was previously used.

 ISSN: 2252-8938
Int J Artif Intell, Vol. 13, No. 1, March 2024: 948-960
950
The result in [8], image processing techniques and data exploration techniques are used to calculate
several significant statistical features, such as the Spatial filters, LBP, Wiener filter, GLCM, Gabor filter, DWT,
Haralick, and segmentation-based fractal texture analysis (SFTA) features. The LBP approach yields the best
rate of identification. It is said to be better than other approaches in many ways. Tsai described a method for
examining the link between printed Chinese characters and digital printers in [9], The feature selection decision
fusion and SVM-based classification are utilized. The most significant features are chosen methodically from
the GLCM, DWT, spatial filters, Wiener filter, and Gabor filter. When compared to other methodologies, the
GLCM method has the highest identification accuracy rate. The word in [10], a collection of characteristics for
describing geometric distortions at the text-line level is presented. Experiments on 14 printers discovered that
the proposed approach outperforms the present state-of-the-art geometric distortion method. When working
with a limited training size, it provides significantly higher accuracy. The average classification accuracy of a
classifier trained with one page, one printer, one font, three different fonts, and 14 printers was 98.85%. A
proposed technique in [11], used all printed letters at the same time to find the originating printer from scanned
images of printed documents. An individual classifier is used to classify all printed letters as well as local
texture pattern-based features. The method was tested on a publicly available dataset of 10 printers as well as
a new dataset of 18 printers scanned at 600 and 300 dpi resolution and created in four different fonts. The
reseacher of [12], used a passive technique to identify the document source printer. Key printer noise features
(KPNF), speeded up robust features (SURF), erientated FAST rotated, and BRIEF are some feature extraction
methodologies that have been used. For the classification job, three classification processes are considered: k-
nearest-neighbor (k-NN), random forest, and decision tree. These three classification techniques received the
most votes. The best accuracy of 95.1% was obtained by combining ORB, KPNF, and SURF with an RF
classifier and an adaptive boosting technique.
Printer identification using GLCM is presented in [13]. A feature vector is created by extracting a set
of features from each character for each letter "e" in the document. Following that, a 5-nearest-neighbor (5-
NN) classifier is used to classify each feature vector. With training, this approach is unaffected by font type or
size, although cross-font and cross-size testing yielded mixed results. A separate 5-NN classifier block for each
character would be required to classify a document using all its characters, not only "e"s. The classifier becomes
more complex as a result. Techniques for colour and picture documents produced by inkjet printers must also
be researched. A text-independent method for adequately describing source printers using deep visual Features
has been implemented in [14]. Using transfer learning on a pre-trained CNN, the system detected 1200
documents from 20 dissimilar (13) laser and (7) inkjet printers. Anselmo investigated systems that can examine
discriminant-printing patterns directly from existing data in [15]. This helped him to eliminate any prior
assumptions about the printing artifact that distinguish each printer. The experimental results proved that the
system is robust to noisy data and outperforms its counterparts. A unique approach based on the SURF, oriented
FAST rotated, and BRIEF feature descriptors is proposed in [16]. Classification was achieved through the use
of Random Forest, Naive Bayes, k-NN, and other classifier combinations. The model was able to correctly
classify the questioned documents and assign them to the appropriate printer. Using a combination of Naive
Bayes, k-NN, Random Forest classifiers, a basic majority voting system, and adaptive boosting algorithms, the
accuracy was 86.5%. Based on source printer identification (SPI), a text-independent technique for identifying
document forgeries is proposed [18]. The top, middle, and bottom parts of the image are divided. There are
two feature extraction algorithms used: HOG and LBP. For printer identification, classification methods such
as decision trees, k-NN, SVM, random forests, bagging, and boosting are taken into consideration. The best
classification accuracy of 96% is attained with the AdaBoost classifier.
3. BACKGROUND
This section introduces convolutional neural networks (CNN) and various types of CNNs, along with
their details. It also covers transfer learning and its advantages. Additionally, this section explains the details
of support vector machines (SVM) and the reasons for using them.
3.1. Convolution neural network (CNN)
CNN is an artificial neural network suggested by Lecun et al. [19]. It has become one of the best-
known images and speech recognition tools. Its common weights network structure is similar to the simulation's
existing biological neural network. This feature can also decrease the network model's complexity and the
number of parameters. CNN can directly use the image as the input, avoiding the traditional identification
method of complex feature extraction and data reconstruction. Table 1 shows details of different types of CNN
models. Convolutional, activation, pooling, and fully connected layers make up most of a typical CNN, as seen
in Figure 1. To extract features from the input images, convolutional layers of CNN are mostly utilized.
Activation layers are used to enable nonlinear mapping, which enhances feature maps' capacity for expression.

951
Typically, as the images are down sampled, the pooling layer reduces the dimension of the features. The fully
connected layers are typically put at the top, near the output layers. Fully connected layers are employed as the
classifier and final output layers for classification tasks [20].
Table 1. Different types of CNN models
Name Network Year Depth Input size Parameters
AlexNet [21] 2012 8 227*227*3 61M
VGGNet-16 [22]
VGGNet-19
2014 16
19
224*224*3 138M
144M
GoogleNet [23] 2014 22 224*224*3 7M
ResNet-18 [24]
ResNet-50
ResNet-101
2016 18
50
101
227*227*3 11.7M
25.6M
44.6M
Inceptionv3 [25] 2016 48 299*229*3 23.9M
Squeeze Net [26] 2016 18 227*227*3 1.24M
XceptionNet [27] 2017 71 299*299*3 22.9M
DarkNet-19 [28]
DarkNet-53
2017 19
53
256*256*3 20.8M
41.6M
ShuffleNet [29] 2018 50 224*224*3 1.4M
Figure 1. The general architecture of CNN [20]
3.2. Transfer learning (TL)
Deep learning's success is also strongly connected to large amounts of data, which means that a lack
of training data can seriously affect the performance of deep learning models. Transfer learning was introduced
to resolve this problem. Transfer learning (TL) is often employed when a new dataset is smaller than the
previous dataset used to train the learned model [30]. Although it has many advantages, the key ones include
reducing training time, improving neural network performance, and not requiring a lot of data. Transfer
learning can be done in two ways: either through feature extraction or fine-tuning. When executing feature
extraction, the pre-trained network is viewed as an arbitrary feature extractor. This allows the input image to
propagate forward, stopping at the pre-specified layer, taking the outputs of that layer as our features, then
applying machine learning classifiers such as SVM. In contrast to feature extraction, when fine-tuning is carried
out, a new fully connected head is constructed and added to the base architecture. The new fully connected
(FC) layer head is randomly set. (Just like any other layer in a new network) and connected to the body of the
original network. Using fine-tuning, we may use networks that have already been trained to distinguish classes
that weren't initially trained. Figure 2 depicts the transfer learning fine-tuning process.
3.3. Support vector machine (SVM)
The SVM classifier is commonly used in machine learning algorithms for binary classification [5].
SVM is a crucial machine learning method that is primarily employed for multi-class classification problems.
Regardless of their dimensional spaces, it has concentrated on classifying features that have been retrieved
beforehand. The best linear decision surface is produced by minimizing training feature vectors. The feature
vectors are designed in the high-dimensional feature space. The SVM determines the extreme margin
hyperplane to partition the various class feature vectors. The performance of the classification is enhanced if
the margin between the vectors is large.

 ISSN: 2252-8938
952
Figure 2. The transfer learning fine-tuning process
4. THE PROPOSED ALGORITHM
Every printer has its own set of fingerprints. On every printed page, there are some fingerprints left
by the printer. These fingerprints are a printer's distinguishing feature. This research provides an algorithm for
identifying the source printer and categorizing the questioned document into one of the printer types. Deep
learning requires a large database to achieve higher accuracy. Therefore, most current research divides scanned
documents into characters, words, or patches to enlarge datasets. Up to our knowledge, the proposed algorithm
is the first research to apply deep learning with limited datasets and without dividing the scanned document
into characters, words, or patches utilizing CNN models' transfer learning. Three different CNN approaches
are employed to determine the types of printers with small datasets in this research. First, a transfer learning of
pre-trained CNN models for feature extractions is employed. It classifies the result using an SVM classifier.
Second, the 2D-CNN is used to generate a trained model from scratch on a given dataset [31], [32]. Then the
pre-trained model is used for feature extraction and classification with SVM. The third approach fine-tunes the
pre-trained CNN models by using transfer learning, which means that a new fully connected layer replaces the
last fully connected or learnable layer of a CNN model with the number of classes in the datasets.
4.1. Pre-processing
The pre-processing phase is utilized in the training and testing phases. For the pre-processing step,
there are three methods: Histogram equalization (HE), gamma correction, and resizing. Histogram equalization
(HE) [33] helps normalize image grey-scale values and improve brightness discrimination between foreground
and background images. The primary objective of the HE is to generate an image with a consistent distribution
throughout the entire brightness scale by using the cumulative density function of the image as a transfer
function [34]. The histogram function is written as (1),
𝐻(𝑓) =
𝐶(𝑓)−𝐶(𝑓)𝑚𝑖𝑛
(𝑊×𝐻)−𝐶(𝑓)𝑚𝑖𝑛
× (𝐺𝐿 − 1) (1)
where, 𝐻(𝑓) denotes the histogram function of the image, 𝐶(𝑓) indicates the cumulative function,
𝐶(𝑓)𝑚𝑖𝑛 signifies the least non-zero value of the cumulative distribution function, W × H gives the image's
number of pixels, and GL defines the number of grey levels utilized. Gamma correction is a nonlinear process
used to manage an image's overall brightness. By translating the values of the input intensity image to new
values, it improves the image's contrast. The Gamma is derived by (2),
𝑁𝑖 = 𝐺(𝛼) × (𝑂𝑖)
1
𝛾 (2)
where, 𝑁𝑖 indicates the new intensity value, Oi represents the old intensity value, G(α) identifies the gray stretch
parameter utilized to linearly scale the outcome on the image of [0, 255], and λ signifies the positive constant.
Gamma can be any value between 0, and infinite mapping is linear when Gamma is 1 (the default value).
Gamma is weighted toward higher (brighter) output values when it is less than 1. The mapping is weighted
toward lower (darker) output values if Gamma is greater than 1 as shown in Figure 3. Finally, the input image
is resized to match each model's input size because each CNN model has an input size. The histogram
equalization and gamma correction of the image are shown in Figure 4.

953
Figure 3. Plots showing three different gamma correction settings
(a)
(b)
Figure 4. Comparison of (a) original image and (b) output image after applying histogram equalization and
gamma correction
4.2. Using transfer learning via feature extraction and SVM classifier
This section proposes an approach for identifying the source printer. Instead of constructing and
training a deep learning model from scratch, we used transfer learning to transfer the learning abilities to our
application. This approach extracted features using 13 well-known pre-trained CNN models (AlexNet, VGG-
16, VGG-19, GoogleNet, DarkNet-19, DarkNet-53, ResNet-18, ResNet-50, ResNet101, SqueezeNet,
XceptionNet, shuffleNet, inceptionv3). SVM is used to classify the resulting features. The model that uses
transfer learning via feature extraction to extract features and classify them using an SVM classifier is shown
in Figure 5.
4.3. Source printer identification based on the 2D-CNN model
This section proposes the 2D-CNN approach for identifying source printers. The method was split
into two parts: feature extraction using a 2D-CNN model and a source printer predictor using an SVM.
Figure 6 depicts the identification procedures: all the scanned documents of the datasets are pre-processed, as
described in section 4.1. After all the documents have been pre-processed, the next step is applying the 2D-
CNN model. A multi-layer neural network is made up of various combinations of convolutional, rectified linear
unit (ReLU), and pooling layers. Train the system using samples, and it will learn the features of images from
various printers. SVM classifiers replaced SoftMax. As a result, the SVM is trained using the outputs from the
preceding layer (Layer 9). The extracted features and the source printer are then identified from the testing
images. After many experiments, the parameters listed in Table 2 were obtained. These parameters were used
to modify the proposed 2D-CNN architecture.

 ISSN: 2252-8938
954
Figure 5. Using the transfer learning approach via extracting features and classifying using an SVM classifier
Figure 6. The architecture of the proposed 2D-CNN
Table 2. 2D-CNN design model architecture
Layer Type Filters Kernel Output Size
Size stride Padding
1 Convolution1 /RELU 32 5 × 5 1 2 256 ×256 ×32
2 Maxpool_1 - 2 ×2 2 0 128 ×128 ×32
3 Convolution 2/RELU 32 5×5 1 2 128×128×32
4 Maxpool_2 - 2×2 2 0 64×64×32
6 Maxpool_3 - 2×2 2 0 32×32×64
8 Average pool_1 2×2 2 0 16×16×64
9 Fully Connected (FC) - - - - 1×1×20
10 SoftMax - - - - 1×1×20
11 Class Output - - - - 1×1×20
4.4. Using transfer learning via fine-tuning
Transfer learning, which has been used in many fields, allows for the transfer of information from one
domain to another [35]. Fine-tuning is a process for training a new ConvNet by transferring the weights of pre-
trained ConvNets (AlexNet, VGG-16, VGG-19, GoogleNet, DarkNet-19, DarkNet-53, ResNet-18, ResNet-50,
ResNet101, SqueezeNet, XceptionNet, shuffleNet, inceptionv3), and it has been effectively employed in a
variety of applications. All pre-trained ConvNets' last FC layer neurons are changed to match the number of
classes in the current classification task. Transfer learning via fine-tuning We're working on a 20-class
classification task. We reduce the output layer of networks from 1,000 to 20 to fit our dataset of 20 printer
classes. The learning and classification layers that exchange in each pre-trained CNN is shown in Table 3.
Stochastic Gradient Descent (SGD) is applied to fine-tune the network utilizing error backpropagation with a
tiny learning rate of 0.0001. Figure 7 depicts a model that employs transfer learning via fine-tuning.

955
Figure 7. Using the transfer learning approach via fine-tuning
Table3. The learning and classification layers that are exchanged in each pre-trained CNN
Network Learning layer replaced ClassLayer replaced
AlexNet Fc8 Output
ResNet-18 Fc1000 Classification Layer predictions
ResNet-50 Fc1000 ClassificationLayer_fc1000
ResNet-101 Fc1000 ClassificationLayer_predictions
GoogleNet loss3-classifier Output
VGG-16 Fc8 Output
VGG-19 Fc8 Output
DarkNet-19 conv19 Output
DarkNet-53 conv53 Output
SqueezeNet conv10 ClassificationLayer_predictions
XceptionNet Predictions ClassificationLayer_predictions
Inceptionv3 Predictions ClassificationLayer_predictions
ShuffleNet node_202 ClassificationLayer_node_203
5. EXPERIMENTAL SETUP AND DATASET
The proposed approaches were implemented in MATLAB R2021b and tested on a DELL PC with the
following configuration: Intel(R) Core (TM) i7-11800H@2.30 GHz, 6 GHz GPU, 16 GB RAM, 64-bit
Windows 11. Several experiments were performed to evaluate the performance of the proposed approaches.
The datasets used to train and test the suggested approaches are described in section 5.1. Section 5.2 describes
the experiment's setup, whereas section 5.3 describes the evaluation measures. The fourth subsection introduces
a discussion of the results. Finally, a comparison of the three approaches to other methods is given and
comparison with other methods.
5.1. Dataset’s description
Khanna et al.'s public dataset [36] is used to test the proposed approach. The documents in this dataset
were printed on 13 laser printers and 7 inkjet printers. Each printer is given a total of 60 documents to consider.
A printer's documents are all one- of-a-kind. The dataset contains documents from three categories: contracts,
invoices, and scientific papers. The printer model's datasets used in this study are listed in Table 4.
5.2. Evaluation metrics
The performance of the proposed algorithm is assessed using a variety of evaluation metrics, including
confusion matrix, accuracy, recall, precision, and F-measure metrics [37]–[39]. The confusion matrix is a cross
table that records the number of occurrences between two rates, the actual classification, and the predicted
classification. The rows show the actual classification, whilst the columns provide a model prediction. The
classes are mentioned in the same order in the rows as in the columns. As a result, the main diagonal is where
the correctly classified items are placed. Table 5 displays metrics for classification evaluations, where TP: true
positive, FN: false negative, FP: false positive, and TN: true negative.

 ISSN: 2252-8938
956
Table 4. Printer models in the dataset
Dataset Printer type Printer # Documents Abbreviations
01 Ink Officejet 5610 60 P_1
02 Laser Samsung CLP 500 60 P_2
03 Laser Ricoh Aficio MPC2550 60 P_3
04 Laser HP LaserJet 4050 60 P_4
05 Laser OKI C5600 60 P_5
06 Laser HP LaserJet 2200dtn 60 P_6
08 Laser Ricoh Afico Mp6001 60 P_7
11 Ink Epson Stylus Dx 7400 60 P_8
13 Ink Unknown 60 P_9
19 Laser HP Color LaserJet 4650dn 60 P_10
20 Laser Nashuatec DSC 38 Aficio 60 P_11
21 Laser Canon LBP7750 cdb 60 P_12
22 Ink Canon MX850 60 P_13
23 Ink Canon MP630 60 P_14
24 Laser Canon iR C2620 60 P_15
26 Ink Canon MP64D 60 P_16
31 Laser Hp Laserjet 4350 o.4250 60 P_17
32 Ink Unknown 60 P_18
49 Laser Hp Laserjet 5 60 P_19
50 Laser Epson Aculaser C1100 60 P_20
Table 5. Metrics for classification evaluations
5.3. Experimental setup
This section describes the experimental setting and examines the outcomes. Three experiments are
carried out. Evaluating the performance using 2D-CNN with SoftMax and 2D-CNN with SVM, evaluating the
performance using pre-trained models as feature extraction, and evaluating the performance using pre-trained
models as fine-tuning.
5.3.1. Performance of transfer learning via feature extraction
As stated in section 4.2, we use pre-trained CNN as feature extractors in our dataset. When pre-trained
CNN are used as feature extractors, classification is done with an SVM. Table 6 summarizes the accuracy of
feature extraction using pre-trained CNN. VGG-19 reports a maximum classification rate of 83.6%. Table 7
shows the performance of all pre-trained CNN models.
Table 6. The accuracy of feature extraction and classification using pre-trained CNN and SVM before and
after pre-processing
Network Accuracy before pre-processing Accuracy after pre-processing
AlexNet 58.3% 72.8%
ResNet18 43% 72.2%
ResNet50 58.1% 82.8%
ResNet101 59.2% 76.7%
GoogleNet 28.3% 62.5%
VGG-16 68.3% 83.3%
VGG-19 69.2% 83.6%
DarkNet-19 69.4% 66.1%
DarkNet-53 59.2% 78.6%
SqueezeNet 44.7% 73.1%
XceptionNet 47.5% 69.2%
Inceptionv3 46.4% 70.8%
ShuffleNet 43.9% 69.2%
Metrix Formula Evaluation Focus
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑇𝑃
𝑇𝑃 + 𝐹𝑃
Evaluate the positive patterns that are exactly calculated from the total predicted
patterns in a positive class.
𝑅𝑒𝑐𝑎𝑙𝑙 𝑇𝑃
𝑇𝑃 + 𝐹𝑁
Determine the fraction of positive patterns that are perfectly classified
F
− 𝑚𝑒𝑎𝑠𝑢𝑟𝑒
2 × 𝑟𝑒𝑐𝑎𝑙𝑙 × 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
𝑟𝑒𝑐𝑎𝑙𝑙 + 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
Denotes the harmonic mean between recall and precision values
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑁 + 𝐹𝑃
Estimates the proportion of exact predictions to all cases that were believed.

957
Table 7. Performance of pre-trained CNN models as feature extractors
Model Accuracy Recall Precision F1_score
AlexNet 72.8% 72.8% 75% 73.3%
ResNet-18 72.2% 72.2% 74.5% 72.4%
ResNet-50 82.8% 82.8% 83.4% 82.8%
ResNet-101 76.7% 76.7% 78.5% 76.7%
GoogleNet 62.5% 62.5% 65.3% 62.8%
VGG-16 83.3% 83.3% 84.3% 83.4%
VGG-19 83.6% 83.6% 84.2% 83.7%
DarkNet-19 66.1% 66.1% 67.8% 66.5%
DarkNet-53 78.6% 78.6% 80.4% 78.9%
SqueezeNet 73.1% 73.1% 74.7% 73.2%
ShuffleNet 69.2% 69.2% 70.4% 69.3%
Inceptionv3 70.8% 70.8% 72.8% 71.2%
Xception 69.2% 69.2% 70.3% 69.3%
5.3.2. Performance using 2D-CNN with SoftMax and 2D-CNN with SVM
The network architecture comprises four different depths to measure performance: 1, 2, 3, and 4
convolutional layers, with each convolutional layer provided with a ReLU layer and a pooling layer. For
comparison, 7-layer, 10-layer, 13-layer, and 16-layer neural network models were created. The data sets are
divided into 70% training and 30% testing. The network architecture is as follows: The 2D-CNN was trained
up to 20 epochs with a batch size of 10 samples. The root mean square propagation (RMSProp) optimizer with
a learning rate of 0.0001 was used. Each scanned document of samples is submitted for training and
classification to the 7, 10, 13, and 16-layer CNN networks (details in Table 2), with the results displayed in
Table 8. Results show that for 2D-CNN with SoftMax and 2D-CNN with SVM, 10-layer CNNs have the
highest accuracy rate. The table also indicates that 2D-CNN with SVM is more accurate than 2D-CNN with
SoftMax.
Table 8. Performance of 2D_CNN with SoftMax and SVM
2D-CNNwith SoftMax 1_conv (7 layers) 84.2% 84.2% 87.1% 93.3%
2_conv (10 layers) 93.8% 93.8% 94.2% 93.7%
3_conv (13 layers) 90.8% 90.8% 91.3% 90.9%
4_conv (16 layers) 92.1% 92.1% 92.9% 91.9%
2D-CNN with SVM 1_conv (7 layers) 97.5% 97.5% 97.6% 97.5%
2_conv (10 layers) 98.8% 98.8% 98.9% 98.8%
3_conv (13 layers) 98.8% 98.8% 98.9% 98.7%
4_conv (16 layers) 98.3% 98.3% 98.5% 98.4%
5.3.3. Performance of transfer learning via fine-tuning
As stated in section 4.4, we fine-tune pre-trained CNN models in our dataset. This method relied on
deep transfer learning CNN architectures to transfer learning weights, which reduced training time,
mathematical calculations, and hardware resource usage. As shown in Table 3, the pre-trained CNN models
were modified in the last fully connected (The learning) layer and classification layers to meet the number of
classes in the dataset, which consists of 20 classes. The accuracy of fine-tuning pre-trained CNN models is
summarized in Table 9. The maximum classification rate reported by DarkNet-19 is 99.2 %. The performance
of pre-trained CNN models as fine-tuning is shown in Table 10.
Table 9. The accuracy of pre-trained CNN via fine-tuning before and after pre-processing
Network Accuracy before pre-processing Accuracy after pre-processing
AlexNet 80.1% 90%
ResnNet18 87.5% 95.8%
ResNet50 91.7% 96.7%
ResNet101 89.2% 95.8%
GoogleNet 82.5% 90.8%
VGG-16 95.8% 96.7%
VGG-19 95% 96.7%
DarkNet-19 91.7% 99.2%
DarkNet -53 93.3% 97.5%
SqueezeNet 86.7% 95%
XceptionNet 91.7% 98.3%
Inceptionv3 90.8% 98.3%
ShuffleNet 90% 94.2%

 ISSN: 2252-8938
958
Table 10. Performance of pre-trained CNN models as fine-tuning
AlexNet 90% 90% 91.3% 89.6%
ResNet-18 95.8% 95.8% 96.4% 95.7%
ResNet-50 96.7% 96.7% 97.1% 96.6%
ResNet-101 95.8% 95.8% 96.3% 96.6%
GoogleNet 90.8% 90.8% 92.3% 90.8%
VGG-16 97.5% 97.5% 97.7% 97.5%
VGG-19 97.5% 97.5% 97.9% 97.5%
DarkNet-19 99.2% 99.2% 99.3% 99.2%
DarkNet-53 95.5% 95.5% 97.7% 97.5%
SqueezeNet 95% 95% 95.4% 95%
ShuffleNet 94.2% 94.2% 95.2% 93%
Inceptionv3 98.3% 98.3% 98.6% 98.3%
Xception 98.3% 98.3% 98.6% 98.3%
5.3.4. Discussion
A new approach is proposed in this research to determine the printer's source. Although much research
on source printer identification has been proposed, they have all been analysed using distinct datasets and
experimental setups. As previously mentioned, many researchers use isolated characters in a text-dependent
framework for experimental purposes. This is the first paper to use CNNs to identify the source printer without
segmenting the document into characters, words, or patches and with small datasets. For VGG-19, 69.2% and
83.6% recognition rates were achieved using pre-trained CNN models as feature extraction before and after
pre-processing, respectively. Recognition rates of 93.8% and 98.8% were achieved using the (2conv) 2D-CNN
approach with SoftMax and SVM classifiers, respectively. The SVM classifier enhanced the 2D-CNN accuracy
by roughly 5% over the initial configuration. DarkNet-19 achieved recognition rates of 91.7% and 99.2%
utilizing pre-trained CNN models as fine-tuning before and after pre-processing, respectively. When fine-
tuning DarkNet-19, the approach achieved an accuracy rate of 99.2% in several testing, as shown in Table 10.
5.3.5. Comparison with other techniques
This section compares the results of three models with some recent algorithms, including
Elkasrawi et al. [37], CNN [14], KPNF+SURF+ORB [12], SURF and ORB with AdaBoost [16] and
HOG+LBP with AdaBoost [18]. Comparison with related work on the dataset of 20 printers is highlighted in
Figure 8. The proposed approaches using 2D-CNN with SVM transfer learning via fine-tuning DarkNet-19
outperform the previous five algorithms reported in the literature.
Figure 8. Comparison of classification accuracies between the three proposed approaches
and existing methods
6. CONCLUSION
Based on source printer identification (SPI), this research proposes a new approach for identifying
document forgeries. This approach trains three different models on the data set to choose the more accurate
model. In the first model, a transfer learning for 13 pre-trained CNN models is used. These models are used as
feature extractors, and SVM is used as a classifier. In this model, VGG-19 with SVM gives the best result with
83.6.5 % accuracy. In the second model, 2D-CNN was used to train from scratch. Then, this trained model was
utilized for feature extraction rather than SoftMax. We used SVM as a classifier. The model achieved 98.5%
accuracy for 2D-CNN and 93.75% accuracy for 2D-CNN-SVM in the studies. We tried the same 13 pre-trained
models in the third model, but we fine-tuned them by retraining each model and modifying the last fully
connected (The learning) layer. The fine-tuned DarkNet-19 gives the maximum classification rate of 99.2%.
We conclude that transfer learning through fine-tuning is more accurate than the other two models based on

959
the results of the three approaches. The second and third approaches' accuracy is better than five recently
published algorithms that used the same dataset.
REFERENCES
[1] J. A. Lewis, “Forensic document examination defined,” in Forensic Document Examination, Elsevier, 2014, pp. 35–44.
[2] 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, Oct. 2013, doi: 10.1016/j.forsciint.2013.07.017.
[3] P. C. Chu, B. Y. Cai, Y. K. Tsoi, R. Yuen, K. S. Y. Leung, and N. H. Cheung, “Forensic analysis of laser printed ink by X-ray
fluorescence and laser-excited plume fluorescence,” Analytical Chemistry, vol. 85, no. 9, pp. 4311–4315, Apr. 2013, doi:
10.1021/ac400378q.
[4] A. K. Mikkilineni et al., “Printer forensics using SVM techniques,” International Conference on Digital Printing Technologies,
vol. 21, no. 1, pp. 223–226, Jan. 2005, doi: 10.2352/issn.2169-4451.2005.21.1.art00064_1.
[5] J. H. Choi, H. Y. Lee, and H. K. Lee, “Color laser printer forensic based on noisy feature and support vector machine classifier,”
Multimedia Tools and Applications, vol. 67, no. 2, pp. 363–382, Jul. 2013, doi: 10.1007/s11042-011-0835-9.
[6] S. Shang, N. Memon, and X. Kong, “Detecting documents forged by printing and copying,” Eurasip Journal on Advances in Signal
Processing, vol. 2014, no. 1, pp. 1–13, Sep. 2014, doi: 10.1186/1687-6180-2014-140.
[7] M. J. Tsai, J. S. Yin, I. Yuadi, and J. Liu, “Digital forensics of printed source identification for Chinese characters,” Multimedia
Tools and Applications, vol. 73, no. 3, pp. 2129–2155, Sep. 2014, doi: 10.1007/s11042-013-1642-2.
[8] M. J. Tsai and I. Yuadi, “Digital forensics of microscopic images for printed source identification,” Multimedia Tools and
Applications, vol. 77, no. 7, pp. 8729–8758, May 2018, doi: 10.1007/s11042-017-4771-1.
[9] M. J. Tsai, C. L. Hsu, J. S. Yin, and I. Yuadi, “Digital forensics for printed character source identification,” in Proceedings - IEEE
International Conference on Multimedia and Expo, Jul. 2016, vol. 2016-Augus, doi: 10.1109/ICME.2016.7552892.
[10] H. Jain, S. Joshi, G. Gupta, and N. Khanna, “Passive classification of source printer using text-line-level geometric distortion
signatures from scanned images of printed documents,” Multimedia Tools and Applications, vol. 79, no. 11–12, pp. 7377–7400,
Dec. 2020, doi: 10.1007/s11042-019-08508-x.
[11] S. Joshi and N. Khanna, “Single classifier-based passive system for source printer classification using local texture features,” IEEE
Transactions on Information Forensics and Security, vol. 13, no. 7, pp. 1603–1614, Jul. 2018, doi: 10.1109/TIFS.2017.2779441.
[12] S. Gupta and M. Kumar, “Forensic document examination system using boosting and bagging methodologies,” Soft Computing,
vol. 24, no. 7, pp. 5409–5426, Aug. 2020, doi: 10.1007/s00500-019-04297-5.
[13] A. K. Mikkilineni, P.-J. Chiang, G. N. Ali, G. T. C. Chiu, J. P. Allebach, and E. J. Delp III, “Printer identification based on graylevel
co-occurrence features for security and forensic applications,” in Security, Steganography, and Watermarking of Multimedia
Contents VII, Mar. 2005, vol. 5681, p. 430, doi: 10.1117/12.593796.
[14] M. Bibi, A. Hamid, M. Moetesum, and I. Siddiqi, “Document forgery detection using printer source identification-a text-
independent approach,” in 2019 International Conference on Document Analysis and Recognition Workshops, ICDARW 2019, Sep.
2019, vol. 2019-Janua, pp. 7–12, doi: 10.1109/ICDARW.2019.70134.
[15] A. Ferreira et al., “Data-driven feature characterization techniques for laser printer attribution,” IEEE Transactions on Information
Forensics and Security, vol. 12, no. 8, pp. 1860–1873, Aug. 2017, doi: 10.1109/TIFS.2017.2692722.
[16] M. Kumar, S. Gupta, and N. Mohan, “A computational approach for printed document forensics using SURF and ORB features,”
Soft Computing, vol. 24, no. 17, pp. 13197–13208, Jan. 2020, doi: 10.1007/s00500-020-04733-x.
[17] P. J. Chiang et al., “Printer and scanner forensics,” IEEE Signal Processing Magazine, vol. 26, no. 2, pp. 72–83, Mar. 2009, doi:
10.1109/MSP.2008.931082.
[18] N. F. E. Abady, M. Taha, and H. H. Zayed, “Text-independent algorithm for source printer identification based on ensemble
learning,” Computers, Materials and Continua, vol. 73, no. 1, pp. 1417–1436, 2022, doi: 10.32604/cmc.2022.028044.
[19] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of the
IEEE, vol. 86, no. 11, pp. 2278–2323, 1998, doi: 10.1109/5.726791.
[20] X. Yu, J. Wang, Q. Q. Hong, R. Teku, S. H. Wang, and Y. D. Zhang, “Transfer learning for medical images analyses: A survey,”
Neurocomputing, vol. 489, pp. 230–254, Jun. 2022, doi: 10.1016/j.neucom.2021.08.159.
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Advances in
Neural Information Processing Systems, vol. 2, pp. 1097–1105, 2012.
[22] K. S. and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” International Conference on
Learning Representations, 2015, doi: 10.48550/arXiv.1409.1556.
[23] C. Szegedy et al., “Going deeper with convolutions,” Proceedings of the IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, vol. 07-12-June, pp. 1–9, 2015, doi: 10.1109/CVPR.2015.7298594.
[24] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition, vol. 2016-Decem, pp. 770–778, 2016, doi: 10.1109/CVPR.2016.90.
[25] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception Architecture for Computer Vision,”
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2016-Decem, pp. 2818–
2826, 2016, doi: 10.1109/CVPR.2016.308.
[26] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally, and K. Keutzer, “SqueezeNet: AlexNet-level accuracy with 50x
fewer parameters and <0.5MB model size,” 2016.
[27] F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” Proceedings - 30th IEEE Conference on Computer
Vision and Pattern Recognition, CVPR 2017, vol. 2017-Janua, pp. 1800–1807, 2017, doi: 10.1109/CVPR.2017.195.
[28] J. Redmon and A. Farhadi, “YOLO9000: Better, faster, stronger,” Proceedings - 30th IEEE Conference on Computer Vision and
Pattern Recognition, CVPR 2017, vol. 2017-Janua, pp. 6517–6525, 2017, doi: 10.1109/CVPR.2017.690.
[29] X. Zhang, X. Zhou, M. Lin, and J. Sun, “ShuffleNet: An extremely efficient convolutional neural network for mobile devices,” in
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Jun. 2018, pp. 6848–6856,
doi: 10.1109/CVPR.2018.00716.
[30] D. Larsen-Freeman, “Transfer of learning transformed,” Language Learning, vol. 63, no. SUPPL. 1, pp. 107–129, Feb. 2013, doi:
10.1111/j.1467-9922.2012.00740.x.
[31] M. J. Tsai, Y. H. Tao, and I. Yuadi, “Deep learning for printed document source identification,” Signal Processing: Image
Communication, vol. 70, pp. 184–198, Feb. 2019, doi: 10.1016/j.image.2018.09.006.
[32] C. B. Rabah, G. Coatrieux, and R. Abdelfattah, “Automatic source scanner identification using 1D convolutional neural network,”

 ISSN: 2252-8938
960
Multimedia Tools and Applications, vol. 81, no. 16, pp. 22789–22806, Jun. 2022, doi: 10.1007/s11042-021-10973-2.
[33] W. K. Pratt, “Digital image processing,” European Journal of Engineering Education, vol. 19, no. 3, p. 377, Jan. 1994, doi:
10.1080/03043799408928319.
[34] Anjani Suputri Devi D and Satyanarayana Ch, “An efficient facial emotion recognition system using novel deep learning neural
network-regression activation classifier,” Multimedia Tools and Applications, vol. 80, no. 12, pp. 17543–17568, Feb. 2021, doi:
10.1007/s11042-021-10547-2.
[35] W. Tao, M. Al-Amin, H. Chen, M. C. Leu, Z. Yin, and R. Qin, “Real-time assembly operation recognition with fog computing and
transfer learning for human-centered intelligent manufacturing,” Procedia Manufacturing, vol. 48, pp. 926–931, 2020, doi:
10.1016/j.promfg.2020.05.131.
[36] N. Khanna, A. K. Mikkilineni, G. T. C. Chiu, J. P. Allebach, and E. J. Delp, “Scanner identification using sensor pattern noise,” in
Security, Steganography, and Watermarking of Multimedia Contents IX, Feb. 2007, vol. 6505, p. 65051K, doi: 10.1117/12.705837.
[37] S. Elkasrawi and F. Shafait, “Printer identification using supervised learning for document forgery detection,” in Proceedings - 11th
IAPR International Workshop on Document Analysis Systems, DAS 2014, Apr. 2014, pp. 146–150, doi: 10.1109/DAS.2014.48.
[38] Y. Zhang, J. Goh, L. L. Win, and V. Thing, “Image region forgery detection: A deep learning approach,” Cryptology and
Information Security Series, vol. 14, pp. 1–11, 2016, doi: 10.3233/978-1-61499-617-0-1.
[39] M. Grandini, E. Bagli, and G. Visani, “Metrics for multi-class classification: an overview,” 2020.
BIOGRAPHIES OF AUTHORS
Naglaa F. El Abady received the M.Sc. degree in computer science at the
Faculty of Computers and Artificial intelligence, Benha University, Egypt, in 2015. She is
currently an assistant lecturer at the computer science department, Faculty of Computers and
Artificial intelligence, Benha University, Egypt. Currently, she is working on his
Ph.D. degree. Her research interest’s concern: digital forensics (document forgery detection),
security (cryptography), and image processing. She can be contacted at email:
naglaa.fathy@fci.bu.edu.eg.
Hala H. Zayed received her B.Sc. in electrical engineering (with an honor
degree) in 1985, her M.Sc. in 1989, and her Ph.D. in 1995 from Benha University in
electronics engineering. She is now a professor at the Information Technology and
Computer Science faculty, Nile University, Egypt. Her research areas are computer vision,
biometrics, machine learning, and image processing. She can be contacted at email:
hala.zayed@fci.bu.edu.eg or hhelmy@nu.edu.eg.
Mohamed Taha is an Associate Professor at Benha University, Faculty of
Computers and Artificial intelligence, Computer Science Department, Egypt. He received his
M.Sc. degree and his Ph.D. degree in computer science at Ain Shams University, Egypt, in
February 2009 and July 2015. He is the founder and coordinator of the "Networking and
Mobile Technologies" program, Faculty of Computers and Artificial Intelligence, Benha
University. His research interest’s concern: computer vision (object tracking-video
surveillance systems), digital forensics (image forgery detection–document forgery
detection-fake currency detection), image processing (OCR), computer networks (routing
protocols-security), augmented reality, cloud computing, and data mining (association
rules mining-knowledge discovery). Taha has contributed more than 20+ technical papers
in international journals and conferences. He can be contacted at email:
mohamed.taha@fci.bu.edu.eg.

Source printer identification using convolutional neural network and transfer learning approach

More Related Content

Similar to Source printer identification using convolutional neural network and transfer learning approach (20)

More from IAESIJAI (20)

Recently uploaded (20)

Source printer identification using convolutional neural network and transfer learning approach