A new local pooling approach for convolutional neural network: local binary pattern

The pooling layer used in CNN models aims to reduce the resolution of image/feature maps while retaining their distinctive information, reducing computation time and enabling deeper models. Max and average pooling methods are frequently used in CNN models due to their computational efficiency; howev...

Full description

Saved in:
Bibliographic Details
Published in:Multimedia tools and applications 2024-04, Vol.83 (12), p.34137-34151
Main Authors: Ozdemir, Cuneyt, Dogan, Yahya, Kaya, Yilmaz
Format: Article
Language:English
Subjects:
Artificial neural networks
Computer Communication Networks
Computer engineering
Computer Science
Data Structures and Information Theory
Datasets
Efficiency
Feature maps
Learning
Methods
Multimedia
Multimedia Information Systems
Neural networks
Performance evaluation
Pixels
Special Purpose and Application-Based Systems
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The pooling layer used in CNN models aims to reduce the resolution of image/feature maps while retaining their distinctive information, reducing computation time and enabling deeper models. Max and average pooling methods are frequently used in CNN models due to their computational efficiency; however, these methods discard the position information of the pixels. In this study, we proposed an LBP-based pooling method that generates a neighborhood-based output for any pixel, reflecting the correlation between pixels in the local area. Our proposed method reduces information loss since it considers the neighborhood and size of the pixels in the pooling region. Experimental studies were performed on four public datasets to assess the effectiveness of the LBP pooling method. In experimental studies, a toy CNN model and various transfer learning models were utilized in conducting test operations. The proposed method provided improvements of 1.56% for Fashion MNIST, 0.22% for MNIST, 3.95% for CIFAR10, and 5% for CIFAR100 dataset using the toy model. In the experimental studies conducted using the transfer learning model, performance improvements of 6.99 ∓ 0.74 and 8.3 ∓ 0.1 were achieved for CIFAR10 and CIFAR100, respectively. We observed that the proposed method outperforms the commonly used pooling layers in CNN models. Code for this paper can be publicly accessed at: https://github.com/cuneytozdemir/lbppooling
ISSN:1573-7721
1380-7501
1573-7721
DOI:10.1007/s11042-023-17540-x