Classification of invariant image representations using a neural network

Khotanzad, A.; Lu, J.-H.

doi:10.1109/29.56063

Cited by 267 publications

(91 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Therefore, the pattern space with 16 ring features is a 16-dimensional unit cube. The Zernike features up to order five defined in [15] are also extracted and then scaled within the range [0,1] along each dime nsion. Pattern space with Zernike features is 16-dimensional unit cube because; selected features start from second order m oments.…”

Section: Case 1: Accommodation By Expansion Of Hsmentioning

confidence: 99%

Modular general fuzzy hypersphere neural network

Patil¹,

Kulkarni²,

Patil³

et al. 2005

17th IEEE International Conference on Tools With Artificial Intelligence (ICTAI'05)

View full text Add to dashboard Cite

This paper describes Modular General Fuzzy Hypersphere Neural Network (MGFHSNN) with its learning algorithm, which is an extension of General Fuzzy Hypersphere Neural Network (GFHSNN) proposed by Kulkarni, Doye and Sontakke [1] that combines supervised and unsupervised learning in a single algorithm so that it can be used for pure classification, pure clustering and hybrid classific ation/clustering. MGFHSNN offers higher degree of parallelism since each module is exposed to the patterns of only one class and trained without overlap test and removal, unlike in Fuzzy Hypersphere Neural Network (FHSNN) [2], leading to reduction in training time. In proposed algorithm each module captures peculiarity of only one particular class and found superior in terms of generalization and training time with equivalent testing time. Thus, it can be used for voluminous realistic database, where new patterns can be added on fly.

show abstract

Section: Case 1: Accommodation By Expansion Of Hsmentioning

confidence: 99%

Modular general fuzzy hypersphere neural network

Patil¹,

Kulkarni²,

Patil³

et al. 2005

17th IEEE International Conference on Tools With Artificial Intelligence (ICTAI'05)

View full text Add to dashboard Cite

show abstract

“…This is computationally the fastest deterministic algorithm; however, it is also the most susceptible to local minima problems. Following Khotanazad and Lu (1990), this learning procedure can be briefly described as follows. The connection weights Wi are initialized to small random values.…”

Section: The Back-error-propagation Learning Algorithmmentioning

confidence: 99%

“…Following Rumelhart and McClelland (1986), and Khotanazad and Lu (1990), learning can be accomplished by the optimization of a criterion function, where constraint satisfaction can be achieved by estimating the discrepancy between the desired and actual output values, feeding back an error signal layer by layer to the inputs, and then adjusting the interconnection weights in such ways as to modify them in proportion to their contribution to the total mean-square-error. This is computationally the fastest deterministic algorithm; however, it is also the most susceptible to local minima problems.…”

Section: The Back-error-propagation Learning Algorithmmentioning

confidence: 99%

Yield and Discrimination Studies in Stable Continental Regions

Mitchell¹

1991

View full text Add to dashboard Cite

“…For example, the Fourier-Mellin integral is costly to compute and converges only under certain strong conditions [15]. The geometric moments, on the other hand, suffer from a high degree of information redundancy [16] and are sensitive to noise; such problems have been investigated by many researchers, e.g., [17][18][19].…”

Section: Introductionmentioning

confidence: 99%