Generated Moment Invariant Features by Cascaded Neural Network for Pattern Classification

スポンサーリンク

概要

• 論文の詳細を見る

• This paper presents a technique on how to improve the intraclass invariance of moment invariant features of noisy images. Moment invariant features possess poor intraclass invariance in the presence of noise. Instead of using computational techniques of extracting moment invariant features from images, we use a trained feature extractor neural net to generate the second and third order moments. The generated moments are used as inputs to a trained classifier neural net which identifies the class the generated moments belong to. Noiseless and noisy binary images of numerals of 32x32 matrix which have been translated, scaled, and rotated are used to determine the feasibility of the above technique. The zero-order regular moments are used in normalizing the binary images. The quality of generalization of the feature extractor neural net on the intraclass invariance of an image is examined. This is done by requiring each individually generated moments to fall within a tolerance bandwidth for the given pattern to be correctly classified. In addition to this, computed moments of normalized and unnormalized binary images are used as inputs to a single neural net to compare the effectiveness of this technique with respect to using generated moments. Back-propagation learning algorithm is used in the training of neural net. The number of hidden units used in the hidden layer of the classifier neural net is between 10 and 100. The cascaded neural net performs much better than using computed moments as inputs to a single neural net that functions as a classifier.

• 一般社団法人情報処理学会の論文
• 1994-02-15

著者

• Omatu Sigeru

  Department Of Computer And System Sciences College Of Engineering Osaka Prefecture University

• Omatu Sigeru

  Department Of Information Science And Intelligent Systems Faculty Of Engineering University Of Tokus

• Raveendran Paramesran

  Department of Electrical Engineering, University of Malaya

• Raveendran Paramesran

  Department Of Electrical Engineering University Of Malaya

関連論文

• New and Used Bills Classification Using Neural Networks (Special Section on Digital Signal Processing)
• Generated Moment Invariant Features by Cascaded Neural Network for Pattern Classification
• Fatigue level estimation of monetary bills based on frequency band acoustic signals with feature selection by supervised SOM
• Three-Stage Method for Selecting Informative Genes for Cancer Classification
• A Reliable Classification Method for Paper Currency Based on the Non-Linear PCA
• On Reliability of Paper Currency Classifiers Using Neural Networks and PCA

スポンサーリンク