Restricted Boltzmann Machine as Image Pre-processing Method for Deep Neural Classifier

The paper presents a novel approach to image preprocessing for feature extraction that is designed for reduction of dimensionality of the classifier which is in this case the convolutional neural network (CNN). The proposed method uses Restricted Boltzmann Machine(RBM) as an Aggregation Method (AM)...

Full description

Saved in:
Bibliographic Details
Published in2019 First International Conference on Societal Automation (SA) pp. 1 - 5
Main Authors Sobczak, Szymon, Kapela, Rafal
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.09.2019
Subjects
aggregation techniques of feature vectors
Computer architecture
Convolutional neural networks
deep learning
Feature extraction
image local binary descriptors
Image recognition
Neurons
Restricted Boltzmann Machine
Training
Online AccessGet full text

Cover

More Information
Summary:The paper presents a novel approach to image preprocessing for feature extraction that is designed for reduction of dimensionality of the classifier which is in this case the convolutional neural network (CNN). The proposed method uses Restricted Boltzmann Machine(RBM) as an Aggregation Method (AM) for binary feature descriptors. The assumption of this technique is that the RBM is performing an dimension expansion of the feature space. Also the type of the data undergoes the transformation from binary to floating point. The conventional approach in convolutional neural networks uses as an input the image that consists of one (grayscale) or three channels (RGB). The method presented herein allows to have the number of channels configurable, as it depends on the size of the Restricted Boltzmann Machine (RBM). The size of the entire network and its parallel implementation makes the architecture usable in real-time systems with reduced memory size.
DOI:10.1109/SA47457.2019.8938039