Related Experiment Video
Updated: May 24, 2026
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
Investigation of efficient features for image recognition by neural networks
Alexander Goltsev1, Vladimir Gritsenko
1International Research and Training Center of Informational Technologies and Systems of National Academy of Sciences of Ukraine, Pr. Glushkova 40, Kiev 03680, Ukraine. agoltsev@adg.kiev.ua
Summary
This study introduces LiRA-features for handwritten digit recognition, reducing feature numbers from 60,000 to 7,000. A modular assembly neural network demonstrated slightly better recognition capabilities than the LiRA perceptron.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
- Pattern Recognition
Background:
- Handwritten digit recognition is a fundamental task in pattern recognition.
- Effective feature extraction is crucial for accurate image recognition.
- Neural networks are widely used for classification tasks.
Purpose of the Study:
- To investigate effective and simple features (LiRA-features) for handwritten digit recognition.
- To propose a feature selection method for reducing feature dimensionality.
- To compare the performance of two neural network classifiers: LiRA perceptron and modular assembly neural network.
Main Methods:
- Development and application of LiRA-features for image recognition.
- Implementation of a feature selection method analyzing neural network connection weights.
- Training and evaluation of a modified 3-layer perceptron (LiRA) and a modular assembly neural network on the MNIST dataset.
Main Results:
- The proposed feature selection method reduced the number of features from 60,000 to 7,000.
- Feature reduction preserved comparable recognition capability while accelerating computations.
- The modular assembly neural network showed slightly superior recognition capability compared to the LiRA perceptron.
Conclusions:
- LiRA-features offer an effective and simple approach for handwritten digit recognition.
- The proposed feature selection method significantly reduces computational complexity.
- Modular assembly neural networks show promise for improved handwritten digit recognition performance.