Authors: D.T. Mane; U.V. Kulkarni
Addresses: Department of Computer Engineering, Pune Institute of Computer Technology, Pune, Maharashtra, India ' Department of Computer Science and Engineering, S.G.G.S. Institute of Engineering and Technology, Nanded, Maharashtra, India
Abstract: Pattern classification is the approach of designing a method to map the inputs to the matching output classes. A novel fuzzy convolutional neural network (FCNN) is proposed in this paper for recognition of handwritten Marathi numerals. FCNN uses fuzzy set hypersphere as a pattern classifier to map inputs to classes represented by the combination of the fuzzy set hypersphere. Given labelled classes, the model designed proved efficient with 100% accuracy on the training set. The two major factors which improve the learning algorithm of FCNN are: 1) extract the dominant features from numeral image patterns using customised convolutional neural network (CCNN); 2) supervised clustering is used to create a new fuzzy hypersphere based on its distance measurement learning rules of fuzzy hypersphere neural network (FHSNN) and pattern classification done by the fuzzy membership function. Performance evaluation of model is done on Marathi numerals large datasets and its performance is found to be superior to traditional convolutional neural network (CNN) model. The obtained results demonstrate the fact that FCNN learning rules can be used as a useful representation for different classification pattern problems.
Keywords: fuzzy hypersphere neural network; FHSNN; convolutional neural network; CNN; pattern classification; supervised clustering.
International Journal of High Performance Computing and Networking, 2019 Vol.15 No.3/4, pp.158 - 169
Received: 06 Oct 2018
Accepted: 07 Apr 2019
Published online: 18 Mar 2020 *