Features extraction for image identification using computer vision

Venant Niyonkuru 1, *, Sylla Sekou 2 and Jimmy Jackson Sinzinkayo 3

1 Department of Computing and Information System, Kenyatta University, Kenya.
2 Department of Mathematics, Institute for Basic Science, Technology and Innovation, Pan-African University, Kenya.
3 Department of Software Engineering, College of Software, Nankai University, China.

World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

Publication history: Received on 01 June 2025; revised on 12 July 2025; accepted on 14 July 2025

Article DOI: https://doi.org/10.30574/wjarr.2025.27.1.2647

Abstract
This study examines various feature extraction techniques in computer vision, the primary focus of which is on Vision
Transformers (ViTs) and other approaches such as Generative Adversarial Networks (GANs), deep feature models,
traditional approaches (SIFT, SURF, ORB), and non-contrastive and contrastive feature models. Emphasizing ViTs, the
report summarizes their architecture, including patch embedding, positional encoding, and multi-head self-attention
mechanisms with which they overperform conventional convolutional neural networks (CNNs). Experimental results
determine the merits and limitations of both methods and their utilitarian applications in advancing computer vision.

Keywords: Feature Extraction; Positional Embeddings; Self-Attention; Vision Transformers (ViTs)

1. Introduction
Feature extraction is a critical stage in the computer vision domain that is the backbone of transforming raw image data
with high amounts into compact, descriptive representations that enable object detection, image categorization,
segmentation, and scene interpretation. Traditionally, feature extraction methods have developed over time based on
the need for creating descriptors that are invariant to scaling, rotation, lighting, and perspective, but computationally
effective (Dosovitskiy et al, 2020; Jiang, 2009; Lowe,2004; Grill et al, 2020).

Traditional feature extraction techniques like Scale-Invariant Feature Transform (SIFT), Speeded-Up Robust Features
(SURF), and Oriented FAST and Rotated BRIEF (ORB) have been instrumental for initial computer vision systems
(Lowe,2004; Rublee et al, 2011, Morrow, 2000) . These algorithms engineer features from local image properties,
finding keypoints and constructing descriptors to facilitate matching among different images. While resistant in the
majority of scenarios, these hand-crafted features are often prone to difficulty with complexity, scalability, and
sometimes devoid of semantic context.

Deep learning transformed feature extraction by the power to learn hierarchical representations directly from data
without needing hand-designed features. Convolutional Neural Networks (CNNs) emerged as the standard by
leveraging local spatial correlation and shared weights but with the expense of local receptive fields, which limit their
capacity to learn long-range dependencies in images (Dosovitskiy et al, 2020; Ali, et al, 2023; Krizhevsky et al,
2012,Morrow, 2000). Here, Vision Transformers (ViTs) have emerged as a highly promising substitute that brings the
self-attention mechanism of NLP into computer vision (Dosovitskiy et al, 2020). ViTs work by dividing images into fixed-
size patches, flattening them, and linearly embedding them. Positional embeddings help to maintain the spatial
information, and the patch embedding is fed into multi-head self-attention to capture global context (Dosovitskiy et al,
2020, Patwardhan et al, 2023, Montrezol, 2024). This paradigm change helps ViTs capture the relationships of the entire


Corresponding author: Venant Niyonkuru
Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution License 4.0.
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

image and surpass limitations intrinsic to CNNs and delivering superior performance across a wide array of vision
benchmarks.

Also, newer architectures such as Generative Adversarial Network (GAN)-based models and contrastive learning
techniques have added to the list of tools used to learn semantic features from images (Ali et al, 2024, Cao et al, 2018;
Kovács et al, 2023).

These are aimed at learning discriminative and generative representations that are useful over a broad range of tasks
ranging from image generation to self-supervised learning (Grill et al., 2020; Ansar et al., 2024). This study
comprehensively examines these varied feature extraction approaches, demystifying the principle behind Vision
Transformers and their position within the wider computer vision context.

Experimental results clarify their individual strengths, compromises, and practical usability, sketching the outline for
the best feature extraction approaches to use in real-world applications(Purchase, 2012).

2. Related work
Traditional approaches are SIFT (Lowe, 2004), SURF (Bay et al., 2008), and ORB (Rublee et al., 2011), which have served
as standard baseline approaches to image matching and recognition. These approaches rely on handcrafted descriptors
to obtain local features. They operate well in structured or low-variation visual scenes. However, they cannot deal with
scale variation, illumination variation, and occlusion. One of the major breakthroughs as exemplified by the emergence
of deep learning models, particularly Convolutional Neural Networks (CNNs), was when these models learned to learn
end-to-end discriminative hierarchical features from raw images (Krizhevsky et al., 2012). CNNs were more
generalizable on a wide variety of vision tasks and thus remained the standard for a number of years.

In recent times, Vision Transformers (ViTs) have been strong competitors that are based on self-attention mechanisms
for obtaining long-range relations in images (Dosovitskiy et al., 2020). ViTs outperformed CNNs on big-benchmark
benchmarks, particularly when they were trained on very large datasets. Simultaneously, Generative Adversarial
Networks (GANs) have not only been utilized for image synthesis but also for feature extraction, depending on
discriminators for obtaining detailed, high-level features. Furthermore, contrastive learning techniques such as BYOL
(Grill et al., 2020) and SimCLR have enhanced self-supervised feature learning by optimizing the agreement between
multiple copies of an image that are transformed differently.

Recent large-scale surveys (Ali et al., 2023; Patwardhan et al., 2023) cover developments in these architectures,
presenting trends and open questions. However, there are fewer papers providing an explicit comparison of these
different approaches under the same experimental setting. This paper fills this gap by comparing classical descriptors,
CNNs, ViTs, and GAN-based models on an identical setup of popular benchmarks and measures.

3. Methodology

3.1. Vision Transformer (Vits)

3.1.1. Definition and functionality

Vision Transformers (ViTs) are deep learning models that leverage self-attention mechanisms to process image data,
offering improved performance over traditional convolutional neural networks (CNNs) (Dosovitskiy et al, 2020) .

3.1.2. Architecture
ViTs divide an image into fixed-size patches, linearly embed them, and feed them into a transformer encoder. The key
components include:

• Patch Embedding Layer: Converts image patches into token embeddings.

• Positional Encoding: Adds spatial information to tokens.
• Multi-Head Self-Attention: Captures long-range dependencies in an image.
• Feed-Forward Network (FFN): Processes token representations for classification tasks.

1342
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

Figure 1 Transformer Encorder

How ViTs Work

• The image is broken into non-overlapping patches

• Each patch is flattened and subsequently passed through a linear projection.
• The transformer encoder converts the patch embeddings through self-attention.
• classification head produces predictions from the last encoded representation.

3.1.3. Image Patches

The process starts with dividing an image into small, fixed-size patches, and that is a simple transformation step. This
process has a direct analogy in natural language processing (NLP) where a sentence is segmented into individual units
such as words or subword tokens. Just like how every token within a sentence carries contextual meaning, every patch
within an image captures localized visual context. In this analogy, the entire image is taken as a sentence, and its patches
are akin to tokens, which enable transformer-based models originally designed for text to be used on visual data.

Figure 2 Image to Image Patches

Both vision transformers (ViT) and natural language processing (NLP) partition large inputs (i.e., sentences in text or
entire images into smaller ones, e.g., tokens in text or image patches). For instance, processing an entire 224×224 pixel
image directly would entail an impossibly large number of calculations, approximately 2.5 billion comparisons. But by
dividing the very same image into 256 patches, each 14×14 pixels, the computation load of one attention layer becomes
incredibly smaller approximately 9.8 million comparisons.

1343
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

Figure 3 Vision transformers

3.1.4. Linear Projection

Following patch division of the image, each patch is then converted from a 2D array to a 1D vector using a linear
projection, effectively projecting raw pixel information into a set of patch embeddings.

Figure 4 Linear Projection

The role of the linear projection layer is to transform each image patch into a fixed-size vector representation, the aim
being to maintain meaningful relations so visually similar patches produce similar embeddings. This transformation
brings the data into a form compatible with the input format needed by the transformer model. Two further processing
steps remain before these embeddings can be used.

3.1.5. Learnable Embeddings

One of the important features added in widely used transformer models such as BERT is the inclusion of a special
classification token, also known as [CLS]. This token is placed at the beginning of every input sequence and is meant to
capture the sentence-level representation for classification tasks.

1344
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

Figure 5 Bert Tokenizer

There is a unique token, [CLS], in BERT that is added to the beginning of all input sequences. This token is embedded
like any other and passed through the encoder layers of the model. The [CLS] token is special in that it doesn't represent
any specific word of the input it begins as a neutral or uninitialized vector. In addition, during pretraining, this final
output at the [CLS] position is fed as input to a classification layer. This encourages the model to encode information
from the entire sentence into this single vector, learning an effective representation of the input. Vision Transformers
(ViT) do exactly the same thing with a learnable embedding that serves the same purpose as the [CLS] token in BERT,
providing a summary representation for image-level classification tasks.

Figure 6 Transformer Encoder to Linear Projection

3.1.6. Positioning Embedding

Transformers do not have an inherent perception of sequence or spatial arrangement of input tokens or patches.
However, preserving order is important in language, where word reordering can dramatically alter meaning. The same
is true for visual information: when the components of a picture are mixed up, as in a jigsaw puzzle, identification of the
whole picture becomes extremely challenging. This is also true for transformer models, which require an additional
mechanism to understand the relative position of these parts.

To address this, positional embeddings are added. In Vision Transformers (ViT), these are learned and of the same
dimension as the patch embeddings. Following the division of the image into patches and adding the special
classification token, each element is added to its respective positional embedding. These position vectors are also
trained along with the model and can further be fine-tuned later. They gradually come to denote spatial relationships,
usually identical to proximate locations in the grid particularly in the same column or row such as:

1345
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

Figure 7 Embeddings Position

Once positional embeddings are added, the patch embeddings are complete. These enhanced embeddings are then
passed into the Vision Transformer (ViT), and they are processed in the same way as regular tokens in a standard
transformer model

Imprementation

1346
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

The training dataset consists of 60,000 images across 11 unique classes. In order to obtain the equivalent human-
readable labels for these classes, the following steps may be used:

ClassLabel has 11 classes: ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog',.].

Each entry in the dataset contains two features: `img` and `label`. The `img` feature contains a 32x32 pixel image which
is of type PIL and with three color channels of RGB (red, green, blue).

3.1.7. Feature extraction

Before sending images to the Vision Transformer (ViT) model, a feature extractor is used to handle preprocessing. This
involves resizing and normalizing images, converting them into tensors referred to as "pixel_values."

1347
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

The feature extractor may be initialized with the Transformers library of Hugging Face, as shown below:

The feature extractor configuration shows that normalization and resizing are set to true. Normalization is performed
across the three color channels using the mean and standard deviation values stored in "image_mean" and "image_std"
respectively.

Therefore, it is optimal to use an image that is slightly larger than needed, since reducing by a small amount usually
preserves visual quality and avoids introducing visible degradation in image quality.

1348
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

Evaluation and Prediction

The Trainer evaluates during training but we can also quickly do a more qualitative verification (or estimation) by
passing through a single image with the model and feature_extractor.

We will pass the following image:

The picture is of poor visual quality and does not have distinguishing features, so visual categorization based on the
picture is difficult.

However, the label given classifies the subject as a cat. We will now go ahead and test the model's prediction for this
picture.

1349
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

Seems like the model is correct!

That's it for our tour of the Vision Transformer and how to use it with Hugging Face Transformers. It's truly remarkable
how quickly transformers have taken over natural language processing and are now making significant incursions into
computer vision.

Just a few years back, before 2021, it was inconceivable to use transformers anywhere except in NLP. But despite being
branded as "those language models," transformers are today at the core of some of the world's most state-of-the-art
computer vision systems. They're components of next-generation architectures like diffusion models [6, 12, 15], and
even power aspects of Tesla's Full Self Driving feature [14, 13, 15, 6].

As time proceeds, we can anticipate a greater convergence of NLP and vision, with transformers persisting in leading
the way of innovation across both fields.

4. Conclusion
To put it briefly, the explosive growth of feature extraction algorithms is an enormous step forward for computer vision
system competence. Handcrafted solutions such as SIFT, SURF, and ORB set the groundwork by providing stable and
interpretable descriptors that are extremely good in specific instances. However, they do poorly when they are tested
against complex, large-scale, or high-variant visual information.

The introduction of deep learning techniques in the form of Vision Transformers (ViTs) is a revolution in paradigm
shifting. By adopting the self-attention mechanism in examining the picture holistically, ViTs have the ability to take
advantage of intricate relationships tricky to model in isolation using convolutional methods alone. Their ability to
include spatial information and condition image patches alike like tokens from NLP have discovered new potential for
better recognition, classification, and downstream visual tasks with high accuracy and robustness.

Furthermore, the inclusion of transformers in autonomous cars such as Tesla's Full Self Driving, and diffusion models,
showcases the growing real-world practical applicability and adaptability of self-attention-based architectures. The
convergence of ideas in NLP and computer vision holds the potential for a future where multimodal models could
perhaps be easily able to handle all kinds of data.

Though their power is remarkable, Vision Transformers are not without problems, including their extensive training
data needs and computational costs. As research continues, hybrid architectures, training methods, and streamlined
transformer variants are being developed to counter these limitations.

In summary, recognizing the advantages and limitations of different feature extraction techniques ranging from classical
to transformer models is vital in pushing computer vision further. As transformers progressively gain the stage, drawing
inspirations from across fields, the next few years promise to be huge in the realization of leaps that will continue to
close the gaps between perception, reasoning, and cognition in artificial systems.

Compliance with ethical standards

Disclosure of conflict of interest

No conflict of interest to be disclosed.

References
[1] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., ... & Houlsby, N. (2020). An
image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929.[2]
A. Vaswani et al., Attention Is All You Need (2017), NeurIPS
[2] Jiang, X. (2009, August). Feature extraction for image recognition and computer vision. In 2009 2nd IEEE
international conference on computer science and information technology (pp. 1-15). IEEE.
[3] Cao, Y. J., Jia, L. L., Chen, Y. X., Lin, N., Yang, C., Zhang, B., ... & Dai, H. H. (2018). Recent advances of generative
adversarial networks in computer vision. IEEE Access, 7, 14985-15006.

1350
 World Journal of Advanced Research and Reviews, 2025, 27(01), 1341-1351

[4] Ali, M., Ali, M., Hussain, M., & Koundal, D. (2024). Generative adversarial networks (gans) for medical image
processing: Recent advancements. Archives of Computational Methods in Engineering, 1-14.
[5] Stable Diffusion (2022), CompVis GitHub Repo
[6] Morrow, G. (2000). Progress in MRI magnets. IEEE transactions on applied superconductivity, 10(1), 744-751.
[7] Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International journal of computer
vision, 60, 91-110.
[8] Grill, J. B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., ... & Valko, M. (2020). Bootstrap your own
latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33,
21271-21284.
[9] Rublee, E., Rabaud, V., Konolige, K., & Bradski, G. (2011, November). ORB: An efficient alternative to SIFT or SURF.
In 2011 International conference on computer vision (pp. 2564-2571). Ieee.
[10] Ali, A. M., Benjdira, B., Koubaa, A., El-Shafai, W., Khan, Z., & Boulila, W. (2023). Vision transformers in image
restoration: A survey. Sensors, 23(5), 2385.
[11] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural
networks. Advances in neural information processing systems, 25.
[12] Patwardhan, N., Marrone, S., & Sansone, C. (2023). Transformers in the Real World: A survey on NLP applications.
Information, 14 (4), 242.
[13] Kovács, L., Csépányi-Fürjes, L., & Tewabe, W. (2023, October). Transformer Models in Natural Language
Processing. In International Conference Interdisciplinarity in Engineering (pp. 180-193). Cham: Springer Nature
Switzerland.
[14] Ansar, W., Goswami, S., & Chakrabarti, A. (2024). A Survey on Transformers in NLP with Focus on Efficiency. arXiv
preprint arXiv:2406.16893.
[15] Patwardhan, N., Marrone, S., & Sansone, C. (2023). Transformers in the Real World: A survey on NLP applications.
Information, 14 (4), 242.
[16] Montrezol, J. (2024). Transforming Medical Imaging: Advancements for ViT (Master's thesis, Universidade do
Porto (Portugal)).
[17] Chowdhary, K., & Chowdhary, K. R. (2020). Natural language processing. Fundamentals of artificial intelligence,
603-649.
[18] Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International journal of computer
vision, 60, 91-110.
[19] Bay, H., Ess, A., Tuytelaars, T., & Van Gool, L. (2008). Speeded-up robust features (SURF). Computer vision and
image understanding, 110(3), 346-359.
[20] Rublee, E., Rabaud, V., Konolige, K., & Bradski, G. (2011, November). ORB: An efficient alternative to SIFT or SURF.
In 2011 International conference on computer vision (pp. 2564-2571). Ieee.
[21] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural
networks. Advances in neural information processing systems, 25.
[22] Grill, J. B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., ... & Valko, M. (2020). Bootstrap your own
latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33,
21271-21284.
[23] Ali-Bey, A., Chaib-Draa, B., & Giguere, P. (2023). Mixvpr: Feature mixing for visual place recognition.
In Proceedings of the IEEE/CVF winter conference on applications of computer vision (pp. 2998-3007).
[24] Patwardhan, N., Marrone, S., & Sansone, C. (2023). Transformers in the real world: A survey on nlp
applications. Information, 14(4), 242.
[25] Purchase, H. C. (2012). Experimental human-computer interaction: a practical guide with visual examples.
Cambridge University Press.

1351