IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 5, October 2025, pp. 3970~3981

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i5.pp3970-3981  3970

Optimizing nitik batik classification through comparative

analysis of image augmentation

Suprapto1, Meilany Nonsi Tentua2, Ahmad Rizki Maulana1

1
Department of Computer Science and Electronics, Faculty of Mathematics and Natural Sciences, Universitas Gadjah Mada,
Yogyakarta, Indonesia
2
Department of Informatics, Faculty of Science and Technology, Universitas PGRI Yogyakarta, Yogyakarta, Indonesia

Article Info ABSTRACT

Article history: Nitik batik is one of the most intricate and culturally significant motifs in
Yogyakarta's batik tradition, characterized by its complex, geometric dot-
Received Nov 21, 2024 based patterns. The unique challenges of automatically classifying nitik
Revised Jul 4, 2025 batik motifs stem from the high variability within the class and the limited
Accepted Aug 6, 2025 availability of training data. This study investigates how different image data
augmentation techniques can enhance the performance of a random forest
classifier for nitik batik motifs. Techniques such as geometric
Keywords: transformations (flip, rotate, and scaling), intensity transformations (cut-out,
grid mask, and random erasing), non-instance level augmentation (pairing
DCGAN samples), and unconditional image generation (deep convolutional
Geometric transformation generative adversarial network (DCGAN)) were used to expand the dataset
Image data augmentation and improve the model's ability to generalize. The results show that specific
Intensity transformation techniques, notably flip, cut-out, and DCGAN, significantly improved
Nitik batik motif classification accuracy, with flip achieving the highest accuracy
improvement of 20.20%, followed by cut-out at 19.27% and DCGAN at
16.25%. Moreover, DCGAN demonstrated the lowest standard deviation
(0.78%), indicating high stability and robustness in classification
performance across multiple validation folds. These findings suggest that
augmentation techniques effectively improve classification accuracy and
enhance the model's ability to generalize from limited and complex datasets.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Suprapto
Department of Computer Science and Electronics, Faculty of Mathematics and Natural Sciences
Universitas Gadjah Mada
North Sekip, Bulak Sumur, Yogyakarta 55281, Indonesia
Email: sprapto@[Link]

1. INTRODUCTION
Batik is Indonesia’s rich and diverse cultural heritage, with motifs that reflect regional identities and
high aesthetic values [1]. Automatic classification of batik motifs is vital in cultural preservation, collection
management, and developing image-based applications related to the creative industry [2]. One of the oldest
batik motifs typical of Yogyakarta is nitik batik. Nitik batik motifs are complex motifs consisting of
thousands of dots arranged and measured in such a way as to form geometric spaces, angles, and fields [3].
The classification of nitik batik motifs faces challenges, including high intraclass variation, pattern
complexity, and limited data available for model training [4].
In computer vision, image classification techniques have overgrown owing to advances in machine
learning algorithms and the availability of extensive data. Some batik classification modeling has been done
using machine learning algorithms such as k-nearest neighbors (KNN) [5], support vector machines (SVM) [6],

Journal homepage: [Link]

Int J Artif Intell ISSN: 2252-8938  3971

backpropagation neural networks (BNN) [7], and decision trees [8]. Meanwhile, classification using deep
learning algorithms for batik modeling mainly uses convolutional neural networks (CNN) [9]. The
performance of classification models is highly dependent on the quality and quantity of data used during the
training process [10], including batik classification modeling. Image data augmentation is an effective
method to improve model performance [11]. Image data augmentation manipulates the original data to
produce new variations that can enrich the dataset. The model can learn better and generalize more to data
that has never been seen before [12].
Augmentation techniques can be categorized into several groups, including geometric
transformation, intensity transformation, non-instance level augmentation, and unconditioned image
generation [13]. Geometric transformation [14] and intensity transformation [15] are often used in
augmentation techniques in batik classification modeling. Both augmentation techniques can increase the
invariance of the model to changes in position. They can help the model deal with lighting and texture
variations in the original image. However, unconditional image generation augmentation techniques have yet
to be used for data addition in classification modeling, especially nitik batik. Unconditional image generation
techniques, such as deep convolutional generative adversarial network (DCGAN), only reproduce new
images to produce previously unknown batik patterns [16], [17].
This study aims to fill this gap by conducting a comparative study of the augmentation methods and
analyzing their impact on the accuracy and robustness of the nitik batik motif classification model. The use of
public datasets on nitik batik cloth motifs provides an opportunity to comprehensively evaluate how these
augmentation techniques affect the performance of the classification model. The goal is to understand better
how these techniques can enhance the effectiveness and efficiency of the nitik batik classification system.
Our contribution is summarized as follows:
‒ The study introduces and compares the effectiveness of multiple image augmentation techniques—
geometric transformations, intensity transformations, non-instance level augmentation, and
unconditioned image generation—for improving nitik batik classification. By exploring a wide range of
augmentation methods, the study provides a comprehensive understanding of how different techniques
impact model accuracy and stability.
‒ By applying the random forest classifier in combination with augmentation techniques, the study
demonstrates significant improvements in classification accuracy, especially through techniques like
flip, cut-out, and DCGAN.
‒ The study emphasizes not only the accuracy but also the stability of the model, as measured by the
standard deviation of cross-validation accuracy. The findings show that certain augmentation
techniques, particularly DCGAN, provide high stability, which is crucial for deploying reliable
classification models in practice.

2. RELATED PAPER
Data augmentation is essential in overcoming dataset limitations by artificially expanding the
training data, while various modeling techniques have emerged to improve image classification
performance. This review examines recent research's role in data augmentation and batik modeling
strategies. Across the studies reviewed, data augmentation geometric transformations techniques were
frequently used for classification model of batik, such as flipping [14], [18], [19], rotation [20], scaling
[21], [22], shearing [23], and noise injection [24] to improve model generalization and reduce overfitting.
More advanced methods, such as random erasing data [25] and brightness modulation [15], [26], are
implemented to input data variations.
Various classification methods have been applied to automate the recognition of batik motifs, each
addressing the unique challenges posed by the complex and highly detailed patterns in batik fabrics.
Traditional machine learning techniques, such as SVM [27], KNN [28], BNN [29], and decision tree [8] have
been widely used for batik classification due to their simplicity and effectiveness in handling small-scale
datasets. More recently, deep learning approaches, particularly CNN, have gained prominence in batik
classification tasks [2], [9], [30]. CNNs automate the feature extraction, learning intricate patterns directly
from raw images, making them highly effective for complex batik motifs, including nitik batik [31]. The
ability of CNNs to capture multi-level features, from edges to textures, has significantly improved the
accuracy and efficiency of batik motif classification. However, CNNs often require large amounts of labeled
data, which can be a limitation for batik datasets.
In addition to CNNs, random forest have been employed for Surakarta batik fabric classification due
to its robustness and ability to handle small, imbalanced datasets [32]. Random forest creates an ensemble of
decision trees, each trained on different parts of the data, allowing the model to capture various features from
batik motifs and providing stable predictions even with limited data. The existing literatures show that image
data augmentation significantly impacts the performance of classification models, especially in the context of
Optimizing nitik batik classification through comparative analysis of image augmentation (Suprapto)
3972  ISSN: 2252-8938

limited or complex datasets. This study aims to compare various augmentation techniques in classifying nitik
batik motifs and provide new contributions to developing more effective classification methods for
culture-based applications.

3. METHOD
The research uses a comparative approach. Several augmentation techniques are applied to the nitik
batik motif dataset, and the results are compared based on classification performance metrics. Figure 1 shows
the stages of the research methodology that were conducted in the research.

Figure 1. The conducted research methodology

3.1. Data collection

The public dataset contains images of nitik batik motifs [33]. This dataset consists of 240 images,
consisting of 60 nitik batik motifs (equal number of images per category). Each image is 512×512 pixels in
size. Figure 2 shows a sample of nitik batik motifs in the dataset.

Figure 2. A sample of nitik batik

3.2. Augmentation technique

In general, classification model including the classification of nitik batik, applying various data
augmentation techniques is essential for enhancing model’s performance. These methods introduce
variability to the dataset, improving the model’s generalization capabilities. This approach evaluates which

Int J Artif Intell, Vol. 14, No. 5, October 2025: 3970-3981

Int J Artif Intell ISSN: 2252-8938  3973

augmentation technique provides the most effective results for accurately classifying nitik batik motifs while
preserving the essential patterns.
The augmentation techniques were selected based on their ability to address the challenges of nitik
batik classification, including high intraclass variability, complex dot-based patterns, and limited dataset size
(240 images). Geometric transformations (flipping, rotation, scaling, shearing, and translation) were chosen to
enhance model robustness to positional and orientational variations, as these are common in real-world batik
images [14], [18], [19]. Intensity transformations (cut-out, grid mask, hide and seek, and random erasing) were
selected to simulate occlusions and lighting variations, improving generalization to imperfect images
[15], [25], [26]. Cut-out and random erasing, in particular, introduce local disruptions, mimicking real-world
imperfections while preserving overall motif structure. Pairing samples was included to explore non-instance-
level augmentation, combining images to create diverse hybrid patterns, as demonstrated in [34]. DCGAN was
chosen to generate synthetic nitik batik images, addressing dataset limitations by producing high-quality
samples that capture intricate textures [16], [17]. These methods were selected to balance diversity,
robustness, and preservation of nitik batik’s cultural and visual characteristics. Table 1 summarizes the
augmentation techniques and their parameters, with a detailed description of the DCGAN architecture
provided. A detailed description of the DCGAN architecture is provided in Figure 3, illustrating the layer
configurations, dimensions, and data flow for both the generator and discriminator.

Table 1. Augmentation techniques and parameters

Augmentation Technique Description and parameters
Geometric Rotation Random rotation within [-30°, 30°], step size of 10°.
transformations [35] Flipping Horizontal and vertical flipping with a probability of 0.5.
Scaling Random scaling between 0.8× and 1.2× of original image size.
Shearing Shear angle range of [-15°, 15°].
Translation Random shifts along X and Y axes within [-10%, 10%] of image dimensions.
Intensity Cut-out Random removal of 1–3 square patches (50×50 pixels) per image.
transformations [36] Grid mask Grid of 10×10 pixel blocks, 50% probability of masking each block.
Hide and seek Randomly hide 16×16 pixel patches, covering 20% of the image.
Random erasing data Erase rectangular regions (10%–20% of image area) filled with random noise.
Non instance level Pairing samples Combine two images with a blending ratio of 0.5 (equal contribution).
augmentation [34]
Unconditional image DCGAN The generator takes a 100-dimensional noise vector and upsamples it through
generation [37] a series of transposed convolutional layers to produce 512×512 grayscale
images, matching the preprocessed nitik batik dataset. The discriminator
evaluates whether input images are real (from the 240-image dataset) or
synthetic. The model was trained for 200 epochs with a batch size of 32,
using the Adam optimizer (learning rate 0.0002, β1 =0.5, β2 =0.999) and
binary cross-entropy loss.

Figure 3. DCGAN architecture for generating synthetic nitik batik images

The DCGAN was implemented using PyTorch, following framework proposed by Radford et al. [38].
Training was conducted on the original 240 nitik batik images, generating 720 synthetic images. These
images were visually inspected to ensure they preserved the characteristic dot-based geometric patterns of
nitik batik motifs. The architecture was tuned to capture the intricate textures of nitik batik, with filter sizes
and layer depths adjusted to handle the 512×512 resolution. All augmentation techniques were implemented
Optimizing nitik batik classification through comparative analysis of image augmentation (Suprapto)
3974  ISSN: 2252-8938

using standard libraries. Parameters were selected based on established practices and tuned to balance
diversity and preservation of nitik batik patterns, ensuring the augmented dataset enhances model
generalization while maintaining cultural and visual integrity.

3.3. Pre-processing
The pre-processing used in our research converts the image into grayscale. This is due to the
relatively limited color variations in nitik batik fabrics. By converting the image to grayscale, the focus shifts
from color information, which is not critical in this case, to the intricate motifs and textures that define nitik
batik. This simplification reduces data complexity and enhances the model’s ability to extract meaningful
features related to the patterns and structures in the batik [39], leading to more efficient and accurate
classification results.

3.4. Feature extraction

We use the binarized statistical image features (BSIF) for feature extraction because it is a practical
feature extraction method to capture batik's complex and unique textures and motifs. BSIF uses filters
learned from natural images to extract local statistical features. This makes it very useful for analyzing
complex patterns such as those in nitik batik.
Images were converted to grayscale to focus on texture patterns, as nitik batik motifs are defined by
dot-based structures rather than color. Grayscale images were normalized to [0, 1] for consistent feature
extraction. BSIF filters, inspired by independent component analysis (ICA) [40], were applied to extract
texture features. These pre-trained filters, derived from natural images, use 8-bit, 5×5 patches to capture local
texture patterns, ideal for encoding the edge-like and geometric features of nitik batik. Visual inspection
confirmed that BSIF processing preserved critical dot-based patterns, and the high classification accuracy
validates the dataset’s suitability. The dataset’s balanced structure (16 images per category) and high
resolution ensure sufficient diversity and detail, making it appropriate for texture-based feature extraction
without losing essential motif characteristics.
The formula for BSIF can be described as follows:
i) A local neighborhood patch represents each pixel in a grayscale image.
ii) A set of learned filters 𝑊1 , 𝑊2 , ⋯ , 𝑊𝑛 is applied to these patches.
iii) The output of each filter is binarized using (1):
𝑇
𝑏𝑖 = {1 if 𝑊𝑖 𝑝𝑎𝑡𝑐ℎ ≥ 0 (1)
0 otherwise

iv) The binary responses 𝑏1 , 𝑏2 , ⋯ , 𝑏𝑛 form a binary string for each pixel, which is then used as the feature
descriptor.
These binary codes efficiently capture local textures, which is particularly useful in recognizing the complex
motifs in nitik batik classification.

3.5. Classification model development

Random forest is an effective classification model for nitik batik because it handles complex,
high-dimensional data such as texture patterns. The model works by building multiple decision trees during
training, each analyzing different features of the batik motif. This model builds multiple decision trees and
combines their predictions for the classification of nitik batik. The critical formula used in random forest is [41]:
i) Gini impurity or entropy for classification trees can be seen at (2) and (3):

𝐺𝑖𝑛𝑖(𝐷) = 1 − ∑𝑖=1 𝑝𝑖2 (2)

𝐸𝑛𝑡𝑟𝑜𝑝𝑦(𝐷) = − ∑𝑛𝑖=1 𝑝𝑖 𝑙𝑜𝑔2 (𝑝𝑖 ) (3)

Where 𝑝𝑖 is the proportion of class 𝑖 in the dataset 𝐷.

ii) Prediction: the final prediction for classification is the majority vote from all trees.
Random forest is well-suited to handle noise and variance in the data. It is a great choice for nitik
batik, where classification depends on detailed texture information rather than primary pixel-based data. To
validate the nitik batik classification model, we employ 𝑘-fold cross-validation, emphasizing the importance
of data diversity and reliability. This approach divides the dataset into 𝑘 = 4 folds of equal size. The model is
trained on (𝑘 − 1) folds and tested on the remaining folds. This process repeats 𝑘, ensuring that every sample
is used for training and validation. By averaging the results, this method reduces bias and variance, thus
providing a robust evaluation of the model's performance on different subsets of data.

Int J Artif Intell, Vol. 14, No. 5, October 2025: 3970-3981

Int J Artif Intell ISSN: 2252-8938  3975

3.6. Evaluation
Once the classification model has been developed, it was evaluated for assessing its accuracy to
determine how well the model can classify different batik motifs. An additional important metric is the
standard deviation of the accuracy across multiple cross-validation folds, as it measures the stability of the
model. A lower standard deviation indicates consistent performance across different data subsets, implying
that the model generalizes well. This combination of metrics ensures that the classification model is accurate
and reliable for real-world application.

4. RESULTS AND DISCUSSION

After applying data augmentation techniques on nitik batik images using geometric transformation,
intensity transformation, non-instance level, and unconditional image generation. Figure 4 illustrates
the preprocessing outcomes for representative nitik batik images. Figure 4(a) shows an original RGB image
of the ‘Sekar cengkeh’ motif, Figures 4(b) to 4(f) displays its geometric transformation augmentation,
Figures 4(g) to 4(j) present the intensity transformation, Figure 4(k) depict the non-instance level
augmentation, and Figure 4(l) shows unconditional image generation augmentation.
The results of flipping in Figure 4(b) and rotating in Figure 4(c) augmentations resemble the
original. Both techniques appear to retain the overall pattern of the motif, only changing the orientation and
direction. These transformations maintain the integrity of the original nitik batik design. The motifs that
change the most from the original are the cut-out in Figure 4(g) and random erasing data in Figure 4(j)
augmentations. These augmentations drastically modify the pattern by removing parts of the motif, thus
changing its appearance significantly. Pairing samples in Figure 4(k) also dramatically alters the image by
combining different motifs, creating a hybrid pattern.

(a) (b) (c) (d) (e) (f)

(g) (h) (i) (j) (k) (l)

Figure 4. Sample result of data augmentation on the sekar cengkeh motif: (a) no augmentation, (b) flip,
(c) rotate, (d) scaling, (e) shearing, (f) translation, (g) cut-out, (h) grid mask, (i) hide and seek,
(j) random erasing data, (k) pairing samples, and (l) DCGAN

BSIF extracts texture features by applying learned filters to local image patches. We use the
parameters filter size 15×15 and filter bit 10×10 to specify the texture extraction in detail. The extracted
features are stored in a feature array and saved to a CSV file. Figure 5 shows the binary feature distribution to
help assess the spread and intensity of the features in the form of a histogram. Figure 5(a) shows the binary
feature distribution for original image, Figures 5(b) to 5(f) display shows the binary feature distribution for
geometric transformation augmentation, Figures 5(g) to 5(j) depict the binary feature distribution for intensity
transformation augmentation, Figure 5(k) present the binary feature distribution for non-instance level, and
Figure 5(l) show the binary feature distribution for unconditional image generation.
Figure 5 presents the plot as a stacked histogram representing the distribution of pixel intensity
values or feature values extracted from several images. Overlaid colors indicate different batches of
processed images. The bell-shaped pattern indicates that most values cluster around the center, reflecting a
balanced intensity distribution or concentration of features around a particular central value. The x-axis
represents the range of intensity values that result from the convolution of the image with BSIF filters
(from -4 to 4), indicating how well the texture in the image aligns with the learned filters. The values give
insights into the strength of the texture patterns and their alignment with the filters. The y-axis on the
histograms represents the frequency or count of pixel intensity values or feature responses extracted using the
Optimizing nitik batik classification through comparative analysis of image augmentation (Suprapto)
3976  ISSN: 2252-8938

BSIF method. The differences in y-axis values across the histograms occur because each augmentation
technique alters the distribution and number of features or pixel values the BSIF filter captures.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k) (l)

Figure 5. The binary feature distribution using BSIF: (a) no augmentation, (b) flip, (c) rotate, (d) scaling,
(e) shearing, (f) translation, (g) cut-out, (h) grid mask, (i) hide and seek, (j) random erasing data, (k) pairing
samples, and (l) DCGAN

Each augmentation method impacts the feature extraction process differently. The BSIF features
extracted after flipping in Figure 5(b), the image may appear quite similar to the original (no augmentation),
as flipping does not significantly alter the local texture but changes the orientation. Thus, the extracted
texture features remain consistent. Rotation, scaling, and shearing in Figures 5(c) to 5(e) introduce noticeable
changes to the features due to alterations in orientation and size. Cut-out, grid mask, and hide and seek in
Figures 5(g) to 5(i) lead to sparse or segmented feature distributions due to occlusion. DCGAN in Figure 5(l)

Int J Artif Intell, Vol. 14, No. 5, October 2025: 3970-3981

Int J Artif Intell ISSN: 2252-8938  3977

shows the impact of synthetic data generation on feature representation. Each augmentation method
affects the stability and consistency of the extracted BSIF features, which directly influences the
classification performance.
After extracting features from each augmented nitik batik image using BSIF, we evaluated the impact
of various augmentation techniques on the accuracy and stability of a random forest-based classification
model. Table 2 compares image augmentation techniques applied to the nitik batik classification model. It
focuses on accuracy and stability (standard deviation) across four cross-validation folds.

Table 2. Comparison of image augmentation techniques applied to nitik batik classification

Augmentation Cross validation (%) Average Standard
1 2 3 4 accuracy (%) deviation (%)
No augmentation 80.00 70.00 85.00 75.00 77.50 5.59
Geometric Flip 9.50 99.16 99.58 99.58 97.70 3.01
transformation Rotate 41.25 43.75 42.91 41.66 42.39 0.99
Scaling 69.16 57.91 62.91 65.41 63.85 4.08
Sharing 65.00 35.00 41.66 42.50 46.04 11.32
Translation 74.58 61.66 70.41 71.66 69.58 4.81
Intensity transformation Cut-out 99.58 95.83 95.41 96.25 96.77 1.65
Grid mask 92.50 70.41 71.25 66.25 75.10 10.22
Hide and seek 37.91 14.58 14.16 13.75 20.10 10.28
Random erasing data 94.16 89.58 89.16 83.75 89.16 3.69
Non-instance level Pairing samples 3.33 11.66 13.75 9.16 9.47 3.90
Unconditional image DCGAN 93.75 92.50 94.58 94.16 93.75 0.78
generation

The evaluation of image augmentation techniques on the nitik batik 960 dataset revealed significant
variations in classification performance using a random forest model, as shown in Table 2. Without
augmentation, the model achieved an accuracy of 77.50% with a standard deviation of 5.59%, indicating
reasonable performance but higher variance across folds. In contrast, augmentation methods like flipping
(97.70% accuracy, 3.01% standard deviation), cut-out (96.77% accuracy, 1.65% standard deviation),
and DCGAN (93.75% accuracy, 0.78% standard deviation) substantially improved both accuracy and
stability. These results underscore the importance of augmentation in enhancing model performance on the
limited and complex nitik batik dataset.
The superior performance of flipping, cut-out, and DCGAN can be attributed to their ability to
balance dataset diversity with the preservation of nitik batik’s intricate dot-based patterns. Flipping maintains
geometric motif integrity by altering orientation without disrupting texture, enabling consistent feature
extraction via BSIF. Cut-out introduces controlled occlusions, simulating real-world imperfections like fabric
wear while retaining sufficient pattern information for robust classification. DCGAN, leveraging CNN
architecture, generates high-quality synthetic images that closely mimic nitik motifs, as confirmed by visual
inspection, providing diverse training samples with minimal variance (lowest standard deviation of 0.78%).
Conversely, methods like pairing samples (9.47% accuracy), hide and seek (20.10% accuracy), and grid
mask (75.10% accuracy) underperformed due to excessive disruption of texture coherence. Pairing samples
creates hybrid patterns by combining unrelated motifs, confusing the classifier, while hide and seek and grid
mask introduce occlusions that obscure critical dot patterns, highlighting the need for augmentation methods
that preserve nitik batik’s unique textures.
Random forest was selected as the classifier due to its alignment with the study’s objective of
evaluating augmentation effects on a small, high-dimensional dataset. Its robustness to BSIF-extracted
texture features and effectiveness on the nitik batik 960 dataset (960 images across 60 motif categories) make
it suitable for capturing intricate dot-based patterns, as evidenced by the high accuracies with flipping and
cut-out. The integration of CNN architecture within DCGAN, which achieved strong accuracy (93.75%) and
the highest stability, further supports random forest’s representativeness. By leveraging CNN-generated
synthetic images, random forest benefits from enhanced dataset diversity without requiring a CNN-based
classifier, maintaining the study’s focus on augmentation impacts. Comparing random forest to other
classifiers like CNNs would necessitate extensive experiments and shift attention away from augmentation
effects, which is beyond this study’s scope.
Despite its strengths, random forest has limitations. Its reliance on handcrafted BSIF features
may not capture multi-level hierarchical patterns as effectively as deep learning models like CNNs,
which learn features directly from raw images. Additionally, its performance is sensitive to augmentation
quality, as seen in the low accuracies of pairing samples and hide and seek, which disrupted motif
coherence. Nevertheless, random forest’s strong performance, interpretability, and alignment with the

Optimizing nitik batik classification through comparative analysis of image augmentation (Suprapto)
3978  ISSN: 2252-8938

study’s goals make it an appropriate choice for assessing the impact of augmentation techniques on
nitik batik classification.
We take the confusion matrix on three classification models that have high performance.
The model was built with data using flipping, cut-out, and DCGAN augmentation. Classification results
are presented in matrix format, where rows represent actual classes, and columns represent predicted
classes. Diagonal elements indicate correct predictions, while off-diagonal elements indicate
misclassification. Figure 6 visualizes the classification performance metrics. Figure 6(a) displays a
confusion matrix for the flipping-augmented dataset. Figure 6(b) presents a confusion matrix for the cut
out-augmented dataset. Figure 6(c) shows a confusion matrix for the DCGAN-augmented dataset. These
results demonstrate that augmentation methods preserving texture coherence significantly enhance
classification performance.

(a) (b) (c)

Figure 6. Confusion matrices of the classification model using different augmentation techniques:
(a) flipping, (b) cut-out, and (c) DCGAN.

Figure 6 shows that each augmentation method has a slightly different effect on the model's
performance, with flip and DCGAN showing the most consistent results. At the same time, cut-out may
introduce more complexity due to image occlusion. The three confusion matrices show that five nitik batik
motifs are incorrectly classified: sekar keben, sekar dangan, gedhangan, sekar pala, and sekar kenanga.
Figure 7 shows the five nitik batik motifs that are incorrectly classified.

Sekar keben Sekar dangan Gedhangan Sekar pala Sekar kenanga

Figure 7. The five nitik batik motifs that were incorrectly classified

The batik motifs displayed (sekar keben, sekar dangan, gedhangan, sekar pala, and sekar kenanga)
may be challenging to predict by the model due to several factors: i) the five nitik batik motifs have similar
geometric shapes or repeating patterns (for example, sekar keben and sekar pala have circular and
symmetrical elements), so they can confuse the model. If the features extracted by the model focus more on
geometric structure than fine details, this can lead to misclassification; ii) subtle texture differences cause the
texture differences to be subtle and not easily captured by the BSIF feature extraction method; and iii) all
motifs have a similar color scheme (mainly dark backgrounds with lighter patterns). Because color
information is less relevant or discarded entirely in grayscale feature extraction methods, the model may need
help differentiating effectively based on texture alone.

Int J Artif Intell, Vol. 14, No. 5, October 2025: 3970-3981

Int J Artif Intell ISSN: 2252-8938  3979

5. CONCLUSION
This research has comprehensively evaluated various image data augmentation techniques in the
context of nitik batik motif classification. By comparing geometric transformations, intensity
transformations, non-instance level augmentation, and unconditional image generation, we identified their
impact on the performance of random forest-based classification models. The results demonstrate that
specific augmentation techniques, namely flipping, cut-out, and DCGAN, significantly enhance the accuracy
of the nitik batik classification model. Flipping achieved the highest accuracy improvement of 20.20%
compared to the baseline model without augmentation, followed by cut-out at 19.27% and DCGAN at
16.25%. Notably, the DCGAN augmentation technique exhibited the highest stability, with a standard
deviation of 0.78%, indicating consistent performance across validation folds. The findings have important
implications for both practical and theoretical domains. Practically, the improved classification accuracy can
be used for the development of automated batik motif recognition systems, supporting cultural preservation
efforts, museum digitization, and applications in the creative industry. Theoretically, this research
demonstrates how different augmentation techniques can overcome the challenges associated with limited
and complex datasets, providing insight into their effectiveness and stability in computer vision tasks. This
research has some limitations. The dataset used is relatively small, consisting of only 240 images, which does
not fully represent the diversity of nitik batik motifs in real-world scenarios. Moreover, the use of grayscale
preprocessing may have overlooked potential color-based features that could improve the classification
performance for specific motifs. Finally, although the random forest classifier proved to be effective, it may
not fully exploit the capabilities of more advanced deep learning models, which could offer further
improvements. For future research, we suggest exploring the incorporation of multiple augmentation
techniques to enhance the model's robustness further. In addition, experimenting with larger and more
diverse datasets, including images with varying illumination or occlusion, may enhance the model's
generalizability. Lastly, evaluating the performance of deep learning architecture along with augmentation
techniques may provide further insights to optimize the classification of nitik batik.

FUNDING INFORMATION
Authors state no funding involved.

AUTHOR CONTRIBUTIONS STATEMENT

This journal uses the Contributor Roles Taxonomy (CRediT) to recognize individual author
contributions, reduce authorship disputes, and facilitate collaboration.

Name of Author C M So Va Fo I R D O E Vi Su P Fu
Suprapto ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Meilany Nonsi Tentua ✓ ✓ ✓ ✓ ✓ ✓
Ahmad Rizki Maulana ✓ ✓ ✓ ✓ ✓ ✓

C : Conceptualization I : Investigation Vi : Visualization

M : Methodology R : Resources Su : Supervision
So : Software D : Data Curation P : Project administration
Va : Validation O : Writing - Original Draft Fu : Funding acquisition
Fo : Formal analysis E : Writing - Review & Editing

CONFLICT OF INTEREST STATEMENT

Authors state no conflict of interest.

DATA AVAILABILITY
The data that support the findings of this study are openly available in Mendeley Data at
[Link]

REFERENCES
[1] A. D. Wibawa, E. A. Wicaksono, S. D. Suryani, and R. Rumadi, “Javanese batik image classification using self-organizing map,” in
2023 International Conference on Computer Science, Information Technology and Engineering (ICCoSITE), 2023, pp. 472–477,
doi: 10.1109/ICCoSITE57641.2023.10127783.

Optimizing nitik batik classification through comparative analysis of image augmentation (Suprapto)
3980  ISSN: 2252-8938

[2] J. Kusanti, E. Noersasongko, M. A. Soeleman, F. Alzami, Purwanto, and Z. A. Hasibuan, “Pre-processing of resize and region of
interest (ROI) to improve the accuracy of batik detection based on VGG-16,” in 2023 10th International Conference on Electrical
Engineering, Computer Science and Informatics (EECSI), 2023, pp. 242–247, doi: 10.1109/EECSI59885.2023.10295666.
[3] A. R. Zuhro, I. K. Sunarya, and W. Nugraheni, “Batik nitik’s existence in the postmodern era,” in Proceedings of the 3rd
International Conference on Arts and Arts Education (ICAAE 2019), 2020, pp. 1–6, doi: 10.2991/assehr.k.200703.001.
[4] A. R. Maulana, Suprapto, and D. A. Tyas, “Classification of batik motifs using multi-texton co-occurrence descriptor and binarized
statistical image features,” Journal of Applied Data Sciences, vol. 6, no. 2, pp. 1367–1378, 2025, doi: 10.47738/jads.v6i2.622.
[5] I. Setyawan, I. K. Timotius, and M. Kalvin, “Automatic batik motifs classification using various combinations of SIFT features
moments and k-nearest neighbor,” in 2015 7th International Conference on Information Technology and Electrical Engineering
(ICITEE), 2015, pp. 269–274, doi: 10.1109/ICITEED.2015.7408954.
[6] A. E. Minarno, Y. Azhar, F. D. S. Sumadi, and Y. Munarko, “A robust batik image classification using multi texton co-occurrence
descriptor and support vector machine,” in 2020 3rd International Conference on Intelligent Autonomous Systems (ICoIAS), 2020,
pp. 51–55, doi: 10.1109/ICoIAS49312.2020.9081833.
[7] C. S. K. Aditya, M. Hani’ah, R. R. Bintana, and N. Suciati, “Batik classification using neural network with gray level
co-occurence matrix and statistical color feature extraction,” in 2015 International Conference on Information & Communication
Technology and Systems (ICTS), 2015, pp. 163–168, doi: 10.1109/ICTS.2015.7379892.
[8] A. Tejawati et al., “Batik pattern classification using decision tree based on color-texture feature,” in 2023 Sixth International
Conference on Vocational Education and Electrical Engineering (ICVEE), 2023, pp. 347–352,
doi: 10.1109/ICVEE59738.2023.10348243.
[9] A. E. Minarno, I. Soesanti, and H. A. Nugroho, “Batik image retrieval using convolutional autoencoder,” in 2024 IEEE 14th
Symposium on Computer Applications & Industrial Electronics (ISCAIE), 2024, pp. 15–20,
doi: 10.1109/ISCAIE61308.2024.10576476.
[10] S. Mohammed et al., “The effects of data quality on machine learning performance on tabular data,” Information Systems,
vol. 132, 2025, doi: 10.1016/[Link].2025.102549.
[11] M. Xu, S. Yoon, A. Fuentes, and D. S. Park, “A comprehensive survey of image augmentation techniques for deep learning,”
Pattern Recognition, vol. 137, 2023, doi: 10.1016/[Link].2023.109347.
[12] A. Mumuni and F. Mumuni, “Data augmentation: a comprehensive survey of modern approaches,” Array, vol. 16, 2022,
doi: 10.1016/[Link].2022.100258.
[13] Y. Zhou, C. Guo, X. Wang, Y. Chang, and Y. Wu, “A survey on data augmentation in large model era,” arXiv-Computer Science,
pp. 1–33, 2024.
[14] I. B. K. Manuaba and V. J. Basiroen, “Comparison study of image augmentation on modified CNN architecture for Indonesian
Lasem-batik’s motifs,” ICIC Express Letters, vol. 17, no. 4, pp. 409–417, 2023, doi: 10.24507/icicel.17.04.409.
[15] M. A. Rasyidi, R. Handayani, and F. Aziz, “Identification of batik making method from images using convolutional neural
network with limited amount of data,” Bulletin of Electrical Engineering and Informatics, vol. 10, no. 3, pp. 1300–1307, 2021,
doi: 10.11591/eei.v10i3.3035.
[16] M. Abdurrahman, N. H. Shabrina, and D. K. Halim, “Generative adversarial network implementation for batik motif synthesis,” in 2019
5th International Conference on New Media Studies (CONMEDIA), 2019, pp. 63–67, doi: 10.1109/CONMEDIA46929.2019.8981834.
[17] A. E. Minarno, M. C. Mustaqim, Y. Azhar, W. A. Kusuma, and Y. Munarko, “Deep convolutional generative adversarial network
application in batik pattern generator,” in 2021 9th International Conference on Information and Communication Technology
(ICoICT), 2021, pp. 54–59, doi: 10.1109/ICoICT52021.2021.9527514.
[18] S. Aras, A. Setyanto, and Rismayani, “Classification of Papuan batik motifs using deep learning and data augmentation,” in 2022 4th
International Conference on Cybernetics and Intelligent System (ICORIS), 2022, pp. 1–5, doi: 10.1109/ICORIS56080.2022.10031320.
[19] D. Trimakno and Kusrini, “Impact of augmentation on batik classification using convolution neural network and k-neareast
neighbor,” in 2021 4th International Conference on Information and Communications Technology (ICOIACT), 2021, pp. 285–289,
doi: 10.1109/ICOIACT53268.2021.9564000.
[20] B. S. Negara, E. Satria, S. Sanjaya, and D. R. D. Santoso, “ResNet-50 for classifying Indonesian batik with data augmentation,” in
2021 International Congress of Advanced Technology and Engineering (ICOTEN), 2021, pp. 1–4, doi:
10.1109/ICOTEN52080.2021.9493488.
[21] I. M. A. Agastya and A. Setyanto, “Classification of Indonesian batik using deep learning techniques and data augmentation,” in
2018 3rd International Conference on Information Technology, Information System and Electrical Engineering (ICITISEE),
2018, pp. 27–31, doi: 10.1109/ICITISEE.2018.8720990.
[22] Farida, R. E. Caraka, T. W. Cenggoro, and B. Pardamean, “Batik parang rusak detection using geometric invariant moment,” in
2018 Indonesian Association for Pattern Recognition International Conference (INAPR), 2018, pp. 71–74, doi:
10.1109/INAPR.2018.8627000.
[23] F. A. Putra et al., “Classification of batik authenticity using convolutional neural network algorithm with transfer learning
method,” in 2021 Sixth International Conference on Informatics and Computing (ICIC), 2021, pp. 1–6,
doi: 10.1109/ICIC54025.2021.9632937.
[24] R. G. Tiwari, A. K. Agarwal, V. Jain, and A. Kumar, “Batik classification in Indonesia: exploring its significance on tourism and
economy,” in 2023 International Conference on Sustaining Heritage: Innovative and Digital Approaches (ICSH), 2023, pp. 119–124,
doi: 10.1109/ICSH57060.2023.10482828.
[25] C. U. Khasanah, E. Utami, and S. Raharjo, “Implementation of data augmentation using convolutional neural network for batik
classification,” in 2020 8th International Conference on Cyber and IT Service Management (CITSM), 2020, pp. 1–5,
doi: 10.1109/CITSM50537.2020.9268890.
[26] M. A. Rasyidi and T. Bariyah, “Batik pattern recognition using convolutional neural network,” Bulletin of Electrical Engineering
and Informatics, vol. 9, no. 4, pp. 1430–1437, 2020, doi: 10.11591/eei.v9i4.2385.
[27] A. Kitipong, W. Rueangsirasak, and R. Chaisricharoen, “Classification system for traditional textile: case study of the batik,” in
2013 13th International Symposium on Communications and Information Technologies (ISCIT), 2013, pp. 767–771,
doi: 10.1109/ISCIT.2013.6645958.
[28] N. Suciati, A. Kridanto, M. F. Naufal, M. Machmud, and A. Y. Wicaksono, “Fast discrete curvelet transform and HSV color
features for batik image clansificotlon,” in 2015 International Conference on Information & Communication Technology and
Systems (ICTS), 2015, pp. 99–104, doi: 10.1109/ICTS.2015.7379879.
[29] N. Suciati, W. A. Pratomo, and D. Purwitasari, “Batik motif classification using color-texture-based feature extraction and
backpropagation neural network,” in 2014 IIAI 3rd International Conference on Advanced Applied Informatics, 2014, pp. 517–521,
doi: 10.1109/IIAI-AAI.2014.108.

Int J Artif Intell, Vol. 14, No. 5, October 2025: 3970-3981

Int J Artif Intell ISSN: 2252-8938  3981

[30] A. E. Minarno, I. Soesanti, and H. A. Nugroho, “A convolutional neural network model for batik image retrieval,” in 2024 IEEE
14th Symposium on Computer Applications & Industrial Electronics (ISCAIE), 2024, pp. 31–36,
doi: 10.1109/ISCAIE61308.2024.10576422.
[31] B. H. Lifindra, D. Herumurti, and A. Yuniarti, “A comparison of VGG architecture convolutional layers in migrating batik style
into fractal shape,” in 2024 International Conference on Smart Computing, IoT and Machine Learning (SIML), 2024, pp. 268–273,
doi: 10.1109/SIML61815.2024.10578183.
[32] A. Fadlil, I. Riadi, and I. J. D. E. P. Putra, “Comparison of machine learning performance using naive Bayes and random forest
methods to classify batik fabric patterns,” Revue d’Intelligence Artificielle, vol. 37, no. 2, pp. 379–385, 2023,
doi: 10.18280/ria.370214.
[33] A. E. Minarno, I. Soesanti, and H. A. Nugroho, “Batik nitik 960 dataset for classification, retrieval, and generator,” Data, vol. 8,
no. 4, 2023, doi: 10.3390/data8040063.
[34] H. Inoue, “Data augmentation by pairing samples for images classification,” arXiv-Computer Science, pp. 1–8, 2018.
[35] C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmentation for deep learning,” Journal of Big Data, vol. 6, no. 1,
2019, doi: 10.1186/s40537-019-0197-0.
[36] H. Naveed, S. Anwar, M. Hayat, K. Javed, and A. Mian, “Survey: image mixing and deleting for data augmentation,”
Engineering Applications of Artificial Intelligence, vol. 131, 2024, doi: 10.1016/[Link].2023.107791.
[37] Y. Lu, D. Chen, E. Olaniyi, and Y. Huang, “Generative adversarial networks (GANs) for image augmentation in agriculture: a
systematic review,” Computers and Electronics in Agriculture, vol. 200, 2022, doi: 10.1016/[Link].2022.107208.
[38] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,”
in 4th International Conference on Learning Representations, ICLR 2016 - Conference Track Proceedings, 2016, pp. 1–16.
[39] W. Liu, N. Lv, J. Wan, L. Wang, and X. Zhou, “Pixel embedding for grayscale medical image classification,” Heliyon, vol. 10,
no. 16, 2024, doi: 10.1016/[Link].2024.e36191.
[40] J. Kannala and E. Rahtu, “BSIF: Binarized statistical image features,” in Proceedings of the 21st International Conference on
Pattern Recognition (ICPR2012), 2012, pp. 1363–1366.
[41] E. Xi, “Image classification and recognition based on deep learning and random forest algorithm,” Wireless Communications and
Mobile Computing, vol. 2022, no. 1, 2022, doi: 10.1155/2022/2013181.

BIOGRAPHIES OF AUTHORS

Suprapto received a B.S. degree in Mathematics from Universitas Gadjah Mada,

Indonesia, in 1987. He received his M.S. degree in computer science from Universitas
Indonesia, Indonesia, and a Doctorate in Computer Science from Universitas Gadjah Mada,
Indonesia, in 2015. From 2001, he was a lecturer at the Department of Computer Science and
Electronics, Universitas Gadjah Mada, Indonesia. Since 2024, he has been a Professor at the
Department of Computer Science and Electronics, Gadjah Mada University, Indonesia. His
research interests include simulation modeling and computational modeling. He is a recipient
of a sandwich program at UI-UMUC, Maryland, USA, in 1990. Since 2021, he has served as
chairman of the Certification Department in the IndoCEISS. Since 2015, he has been
registered as a member of Aptikom and IAENG. He can be contacted at email:
sprapto@[Link].

Meilany Nonsi Tentua received a B.S. degree in Mathematics from Universitas

Gadjah Mada, Indonesia, in 1996. She received a Master’s Degree in Electrical Engineering
from Universitas Gadjah Mada, Indonesia, in 2004, and a Doctorate in Computer Science
from Universitas Gadjah Mada, Indonesia, in 2023. She has been a lecturer at Informatics,
Universitas PGRI Yogyakarta, Indonesia, since 2006. Her research interests include machine
learning, data science, data mining, natural language processing, and artificial intelligence.
She can be contacted at email: meilany@[Link].

Ahmad Rizki Maulana received a B.S. degree in Informatics from Universitas

Negeri Surabaya, Indonesia, in 2021. He is currently pursuing an M.S. degree in Computer
Science at Universitas Gadjah Mada, Indonesia. His research interests include machine
learning, data science, computer vision, and natural language. He can be contacted at email:
ahmadrizkimaulana@[Link].

Optimizing nitik batik classification through comparative analysis of image augmentation (Suprapto)