Image recognition using analog-ART1 architecture augmented with moment-based feature extractor
Introduction
The problems involving numerical calculations are normally solved using various architectures of neural networks by training the samples collected and hence infer some other samples out of training the network as individual sessions to validate the integrity of the system. There are many approaches available to pattern recognition [4] and the most important of them are (i) statistical, (ii) syntactic, and (iii) neural approaches. The Adaptive Resonance Theory (ART) was developed by Carpenter and Grossberg [1], which also includes ART1, a classification for clustering binary vectors and another form ART2 [2] for clustering analog vectors. A learning system must remain adaptive (plastic) in response to significant input yet remain stable to irrelevant inputs and the situation is called stability–plasticity dilemma. The ART attempts to address the stability–plasticity dilemma and the architecture normally considered is ART1 which requires the input vectors to be binary. Here it is proposed to implement a variation of ART1 known as Analog-ART1, where the input vectors need not be binary, they could assume any analog value depends upon the colour chosen during the creation of patterns. This network utilizes the concept of competition between units and this limitation is not necessarily a severe handicap and reduces the computational overhead and architectural redundancy of SFAM [5], [6].
However, Grossberg's ART1 despite its simplicity does not display the characteristic of tolerance to pattern perturbations or noise while processing images. The aim of this investigation is therefore to enhance the pattern recognition capability of ART1 by retaining its simple architecture but by augmenting it with a moment-based feature extractor to enable to display tolerance to pattern perturbances and noise. The feature extractor extracts features from patterns, which are “preprocessed” inputs to ART1. The digital approximations of moment-based invariants have been employed and the approximations invariant to rotation, translation and scaling change. The invariant values are converted to analog values and input to ART1. Thus, the Analog-ART1 is augmented with the above said modified feature extractor has been investigated for its recognition of monochrome and colour images.
In this paper, the architecture of Analog-ART1 has been presented in Section 2. The algorithm of Analog-ART1 architecture is presented in Section 3. The moment-based feature extractor is reviewed in Section 4. The recognition of monochrome and colour images of Analog-ART1 has been presented in Section 5. The results of computer simulations have also been presented in this section.
Section snippets
ART1—a review
The basic feature of the Analog-ART1 architecture is shown in Fig. 1. There are two major subsystems known as attentional subsystem and orienting subsystem. In the two layers of attentional subsystem, some patterns of activity are developed and they are called short-term memory (STM). The weights associated with the bottom up and top own connections between F1 and F2 are called long-term memory (LTM) because they encode information that remains part of net work for an extended period. The
Algorithm analog-ART1()
1. Apply an input vector I to F1 and F1 activities are calculated according to2. Output vector of F1 are calculated as3. Propagate S forward to F2 and calculate the activities according to4. Only the winning node has nonzero output5. Propagate the output from F2 back to F1 and calculate the net inputs from F2 to the F1 unit as6. Calculate maximum activities according to
Feature extractor—moment-based invariants
The moments are extracted features that are derived from raw measurements. In practical imagery, images are subjected to various geometric disturbances or pattern perturbations. It is therefore necessary that features that are invariants to orientations be used for purposes of recognition or classifier. For 2D images moments have been used to achieve Rotation (R), Dilation (D) and Translation (T) invariants and are shown in Table 1 [5], [6]. The normalized moments η is defined as
Recognition of monochrome images
The problem of distinguishing among airplanes, tanks and helicopters from a satellite discussed by Wang et al. [7], [8] is the best suite problem. The experiments conducted are categorized as in Table 2. Fig. 2 illustrates a set of sample nominal patterns (in monochrome) that Analog-ART1 is trained with and the model is tested for a set of seven patterns rotated or scaled or translated or combinations of one or more or all of these along with a noisy image. The moment invariants for all the
Conclusion
With the experiments conducted over monochrome and colour images with the presence and absence of the perturbances and disturbances, We conclude that the advantages of Analog-ART1 are
- •
It learns constantly but learns only significant information.
- •
The information already learnt is not destroyed by new information.
- •
Any input pattern is recalled rapidly if it has been already learnt.
- •
It has the ability to create new categories.
- •
Even though original ART1 was designed to handle binary images with moment
Acknowledgements
The authors express their profound thanks to the Management, Dr. S. Vijayarangan, Principal, PSG College of Technology, Coimbatore 641 004 for providing necessary facilities and constant encouragement to carry out this work.
S. Rajasekaran is a professor of Infrastructure Engineering of PSG College of Technology, Coimbatore, India. He obtained his Ph.D. in Civil Engineering from the University of Alberta, Edmonton, Canada in 1971 and D.Sc (Civil Engineering) from the Bharathiar University, India in October 1999. He was a visiting professor at the University of Alberta, Canada, University of Sydney, Australia and the Alexander von Humboldt Guest Professor at the University of Stuttgart, Gemany. He is a recipient of
References (8)
- et al.
ART2AAn adaptive resonance algorithm for rapid category learning and recognition
Neural Networks
(1991) - et al.
ART2Self-organisation of stable category recognition codes for analog input patterns
Appl. Opt.
(1987) - et al.
Neural Networks Algorithm, Applications and Programming Techniques
(1992) Training free counter propagation neural networks to pattern recognition in fabric defects
Textile Res. J.
(1997)
Cited by (4)
MicroARTMAP for pattern recognition problems
2007, Advances in Engineering SoftwareFine-grained recognition of surface targets with limited data
2020, Electronics (Switzerland)Artificial neural networks: A practical course
2016, Artificial Neural Networks: A Practical Course
S. Rajasekaran is a professor of Infrastructure Engineering of PSG College of Technology, Coimbatore, India. He obtained his Ph.D. in Civil Engineering from the University of Alberta, Edmonton, Canada in 1971 and D.Sc (Civil Engineering) from the Bharathiar University, India in October 1999. He was a visiting professor at the University of Alberta, Canada, University of Sydney, Australia and the Alexander von Humboldt Guest Professor at the University of Stuttgart, Gemany. He is a recipient of the ISTE National Award for his outstanding research work in Engineering and Technology in the year 1991, Tamilnadu Scientist Award by TNCST, Govt. of Tamilnadu in 1996 and NAGADI Award for his book “Finite Element Analysis in Engineering Design” by the Association of Consulting Civil Engineers (ACCE) in 1996 and the Vocational Excellence Award by the Coimbatore West Rotary in Dec 1999 and ISTE Anna University National Award for the Outstanding Academic for the year 1999–2000. Rajasekaran has been the principal investigator for many projects sponsored by AICTE, ARDB, ISRO, BARC, DST and MHRD. He is a fellow of the institution of Engineers, Institute of Valuers and a member of American Society of Civil Engineers. He has published more than 280 research papers in national and international journals and conferences besides 17 books including one on “Neural Networks, Fuzzy Logic and Genetic Algorithms”. He is listed in the American Biographical Research Institute and the Eminent Personalities of India. His specific interests include Finite Element Analysis Boundary Element method, Nonlinear analysis, neural networks, genetic algorithms and fuzzy systems. Recently he has been elected as fellow of National Academy of Engineering.
R. Amal Raj is a Lecturer (Selection grade) in computer Science in the Department of Mathematics and Computer Science at Sri Vasavi College , Erode. He did his Ph.D. programme in Computer Science as a Full time scholar under Faculty Development programme of University Grants Commission under the guidance of Dr. S. Rajasekaran.