Authors

  • Alibek Olimov
    New Uzbekistan University

DOI:

https://doi.org/10.71337/inlibrary.uz.ijai.73085

Abstract

Computer vision has evolved significantly, transitioning from classical techniques based on handcrafted features to deep learning models that learn representations automatically. This article compares classical and deep learning approaches in terms of feature extraction, model complexity, computational efficiency, and application performance. While classical methods like Scale-Invariant Feature Transform (SIFT), Histogram of Oriented Gradients (HOG), and edge detection have been effective in controlled environments, deep learning models, particularly Convolutional Neural Networks (CNNs), have demonstrated superior adaptability in complex real-world scenarios. The discussion highlights their advantages, limitations, and future directions in computer vision research.

 

 

background image

INTERNATIONAL JOURNAL OF ARTIFICIAL INTELLIGENCE

ISSN: 2692-5206, Impact Factor: 12,23

American Academic publishers, volume 05, issue 03,2025

Journal:

https://www.academicpublishers.org/journals/index.php/ijai

page 202

COMPARING CLASSICAL AND DEEP LEARNING APPROACHES

IN COMPUTER VISION

Alibek Olimov Ulugbekovich

New Uzbekistan University Faculty of School computing

student of 4 - course of Software engineering

alibekolimov.info@gmail.com

Abstract:

Computer vision has evolved significantly, transitioning from classical techniques

based on handcrafted features to deep learning models that learn representations automatically.

This article compares classical and deep learning approaches in terms of feature extraction,

model complexity, computational efficiency, and application performance. While classical

methods like Scale-Invariant Feature Transform (SIFT), Histogram of Oriented Gradients (HOG),

and edge detection have been effective in controlled environments, deep learning models,

particularly Convolutional Neural Networks (CNNs), have demonstrated superior adaptability in

complex real-world scenarios. The discussion highlights their advantages, limitations, and future

directions in computer vision research.

Keywords:

Computer vision, classical methods, deep learning, CNN, feature extraction, image

processing, SIFT, HOG, Res Net, AI.

Introduction:

Computer vision (CV) is a subfield of artificial intelligence (AI) that enables

machines to interpret and process visual data, mimicking human vision. Early CV systems

depended on handcrafted features and rule-based algorithms, requiring significant domain

expertise. However, with the advent of deep learning, models can now learn representations

automatically, significantly improving performance in tasks like image classification, object

detection, and segmentation. This article compares classical and deep learning approaches,

highlighting their differences in methodology, efficiency, and real-world applications (1,97). The

discussion aims to provide insights into when classical methods remain viable and when deep

learning offers a clear advantage.

Classical Computer Vision Approaches

.

Classical computer vision primarily relies on

feature engineering and traditional machine learning algorithms to analyze visual data.

Feature Extraction Methods. Feature extraction is crucial in classical computer vision. Some of

the most widely used techniques include:

SIFT (Scale-Invariant Feature Transform) – A technique that detects key points and extracts

invariant descriptors useful for object recognition and image matching.

HOG (Histogram of Oriented Gradients) – Commonly used in pedestrian detection, HOG

captures edge orientations in an image, making it effective for structured objects.

Canny Edge Detection – A popular edge detection algorithm that identifies boundaries in an

image using gradients and non-maximum suppression.


background image

INTERNATIONAL JOURNAL OF ARTIFICIAL INTELLIGENCE

ISSN: 2692-5206, Impact Factor: 12,23

American Academic publishers, volume 05, issue 03,2025

Journal:

https://www.academicpublishers.org/journals/index.php/ijai

page 203

These methods allow machines to extract useful information from images before applying

classification or recognition algorithms.

Machine Learning-Based Classical Methods. Once features are extracted, they are fed into

traditional machine learning models for classification or regression. Popular models include:

Support Vector Machines (SVM) – Works well for binary classification tasks and is often used

in image recognition.

Random Forests – A collection of decision trees that provide robustness against overfitting.
K-Nearest Neighbors (KNN) – A simple but effective algorithm for pattern recognition based

on feature similarity (2,125).

Limitations of Classical Approaches. Despite their usefulness, classical methods have

limitations:
Require manual feature engineering, which demands domain expertise.
Struggle with large and complex datasets due to lack of scalability.
Perform poorly in uncontrolled environments with variations in lighting, perspective, and

occlusion.

Deep Learning Approaches in Computer Vision. Deep learning has revolutionized computer

vision by automating feature extraction and enabling end-to-end learning.

Key Architectures in Deep Learning

Several deep learning architectures have played a crucial role in computer vision:

1.LeNet-5 – One of the earliest CNNs designed for handwritten digit recognition.

2.AlexNet – A deep CNN that won the 2012 ImageNet competition, demonstrating the power

of deep learning.
3.VGGNet – Introduced deeper networks with small 3x3 filters for better feature extraction.
4.ResNet – Introduced residual connections to address the vanishing gradient problem, allowing

the training of very deep networks.

Advantages of Deep Learning

Automatic Feature Learning – Unlike classical methods, CNNs learn feature representations

automatically.

Scalability – Deep learning models can handle large datasets and complex problems efficiently.
Better Generalization – Deep networks perform well in diverse and real-world applications

(3,89).


background image

INTERNATIONAL JOURNAL OF ARTIFICIAL INTELLIGENCE

ISSN: 2692-5206, Impact Factor: 12,23

American Academic publishers, volume 05, issue 03,2025

Journal:

https://www.academicpublishers.org/journals/index.php/ijai

page 204

Limitations of Deep Learning

High Computational Cost – Requires powerful GPUs and extensive computational resources.
Data Dependency – Needs large amounts of labeled data for training.
Interpretability Issues – Deep models are often considered black-box systems, making it

difficult to explain their decisions.

Comparative

Analysis

Applications and Case Studies.

Flower Detection (Example from ML Research)
Classical Approach: SIFT + SVM for feature extraction and classification.
Deep Learning Approach: CNN-based model (e.g., Res Net, Efficient Net) for direct image

classification.

Comparison: CNNs outperform classical approaches in terms of accuracy and robustness.
Other Applications. Medical Imaging: Deep learning surpasses classical methods in detecting

tumors and anomalies in medical scans (4,295).

Autonomous Vehicles: Classical edge detection is useful for lane detection, but CNNs provide

comprehensive object recognition and scene understanding.

Facial Recognition: Classical methods use eigenfaces and SIFT, whereas deep learning

employs CNN-based architectures like Face Net (5, 97).


background image

INTERNATIONAL JOURNAL OF ARTIFICIAL INTELLIGENCE

ISSN: 2692-5206, Impact Factor: 12,23

American Academic publishers, volume 05, issue 03,2025

Journal:

https://www.academicpublishers.org/journals/index.php/ijai

page 205

In this picture visually represents the evolution of Computer vision from Classical method to

Deep learning approaches.

Conclusion:

The evolution of computer vision from classical methods to deep learning has

significantly enhanced the field’s capabilities. Classical approaches, which rely on manually

crafted features and traditional machine learning models, are still valuable for tasks that require

low computational resources and well-defined image structures. These methods perform well in

controlled environments but struggle with scalability and adaptability when applied to large,

diverse datasets (6,345).


background image

INTERNATIONAL JOURNAL OF ARTIFICIAL INTELLIGENCE

ISSN: 2692-5206, Impact Factor: 12,23

American Academic publishers, volume 05, issue 03,2025

Journal:

https://www.academicpublishers.org/journals/index.php/ijai

page 206

On the other hand, deep learning, particularly convolutional neural networks (CNNs), has

revolutionized computer vision by enabling automatic feature learning and significantly

improving performance in complex tasks like image recognition, object detection, and

segmentation. CNNs have demonstrated superior accuracy, generalization, and robustness,

making them the preferred choice for large-scale applications in autonomous driving, healthcare,

and security systems. However, deep learning is not without challenges - it requires vast amounts

of labeled data, high computational power, and remains a black-box model in many cases,

limiting interpretability.

Both classical and deep learning approaches play essential roles in computer vision.

Classical methods remain relevant for tasks requiring lower computational power and well-

defined feature extraction, whereas deep learning dominates large-scale, complex applications

(7,118). Future research should focus on improving deep learning interpretability and reducing

computational costs, making AI-driven vision systems more accessible and efficient.

References:

1. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press 775p.
2. Szeliski, R. (2022). Computer Vision: Algorithms and Applications (2nd ed.), Springer 925p.
3. Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International

Journal of Computer Vision, 60(2), 91-110p.

4. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition.

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),

770-778p.

5. Krizhevsky, A., Sutskever, I., & Hinton, G. (2012). ImageNet classification with deep

convolutional neural networks. Advances in Neural Information Processing Systems

(NeurIPS), 197p.

6. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified,

real-time object detection. Proceedings of the IEEE Conference on Computer Vision and

Pattern Recognition (CVPR), 779-788p.

7. Russakovsky, O., Deng, J., Su, H., et al. (2015). ImageNet large scale visual recognition

challenge. International Journal of Computer Vision, 115, 211-252p.

References

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press 775p.

Szeliski, R. (2022). Computer Vision: Algorithms and Applications (2nd ed.), Springer 925p.

Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2), 91-110p.

He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770-778p.

Krizhevsky, A., Sutskever, I., & Hinton, G. (2012). ImageNet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems (NeurIPS), 197p.

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779-788p.

Russakovsky, O., Deng, J., Su, H., et al. (2015). ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115, 211-252p.