Fukushima's Neocognitron is a hierarchical neural network and has shown fair amount of success in handwritten pattern recognition due to its ability to recognize deformed visual patterns. Some difficulties of implementing Neocognitron are the complexity of the model and its 2-dimensional approach. Recently, RSSIPL developed two new neural networks, called Tricognitronand Shape Cognitron based on Fukushima's Neocognitron so as to achieve shift invariance, scale invariance, deformation invariance and distortion invariance. A software designed to run Tricognitron on PCs has been developed and tested successfully for 35 alphanumeric characters. The Shape-Cognitron jointly developed by National Cheng Kung University in Taiwan is a generalized and modified version of Tricognitronwhich extends the ability of Tricognitronto recognition of shape orientations of objects. Applications of Shape Cognitron include recognition of different shapes of clustered microcalcifications in mammograms, patterns of lung nodules in chest radiographs, geometric features of blood vessels. As a result, Shape Cognition can be used for classification of malignant/benign or normal/abnormal diseases. The idea of Shape Cognitron has been filed for a US patent by UMBC.

Another technique developed in RSSIPL is to use the concept of principal components analysis as a base to design automatic neural network systems for documentation or text recognition and any type of gray-level target pattern recognition including gray-level character pattern recognition, fingerprints verification, vehicle identification, etc. The PCA-based neural network is particularly useful for patterns to be recognized whose size are (1) small in terms of pixel size such as 5x10 or (2) which are gray-level images. In either case, current OCR cannot do well due to small blurring or small noise effects. Both cases could be fatal in applications such as documentation and text analysis (e.g., tax return form or check verification, they are gray-level and also small), target pattern recognition, gray level fingerprint verification; medical imaging (mass detection) and speech recognition. RSSIPL also developed a PCA-based neural network which was successfully applied to the data used for Fire Support Combined Arms Tactics Trainer Program sponsored by US Army. A software package was licensed to AAI company for real time implementation.

Invention Honors

  1. Recognition was among 10 finalists in the April 1993 BFGoodrich Collegiate Inventors Program Competition in USA and was awarded honorable mention.
Go back to Menu

Patent Awarded

  1. Y. Xu, and C.-I Chang, Method and Apparatus for Pattern Recognition, to be issued July, 1998.
Go back to Menu

Disclosure of Invention

  1. Y. Xu and C.-I Chang, Tricognitron: A New Approach to Pattern Recognition for Line-Curve Configurations, November 11, 1992, filed with UMBC.
  2. C.-I Chang, Pattern Recognition Using PCA-Based Neural Networks, March 1, 1996.
  3. C.-I Chang, and B.-C. Hsu, Shape Cognition for Pattern Recognition, 1996, in preparation.
Go back to Menu

Software License

  1. C.-I Chang, Energy-Based Neural Networks for Pattern Recognition, Sept, 1996.
Go back to Menu

Master Theses

  1. H. Gandhi, Methodologies of Handwritten Character Recognition, M.S. Thesis, Department of Electrical Engineering, University of Maryland Baltimore County, MD, May 1994
  2. Y. Xu, Tricognitron: A New Approach to Neocognitron for Handwritten Alphanumeric Character Recognition M.S. Thesis, Department of Electrical Engineering, University of Maryland Baltimore County, MD, August 1993
Go back to Menu

Publications

  1. Y. Xu and Chein-I Chang, "Implementation of 3-D model for Neocognitron," submitted to International Conference on Neural Network, Washington, DC, June, 1996, pp. 794-799.
  2. J.-R. Tsai. P.C. Chung and C.-I Chang, "A sigmoidal radial basis function neural network for function approximation," International Conference on Neural Network, Washington DC, June 3-6, 1996, pp. 496-501.
  3. C.-W. Yang, P.-C. Chung, B.-Chang Hsu and C.-I Chang, "A hierarchical neural model for shape recognition: shape cognitron,"1996 International Symposium on Multi-Technology Information Processing, Kaohsiung, Taiwan, ROC, 1996.
  4. R.H. Baran, H. Ko and C.-I Chang, "Signal detectability with neural network," Proc. Asia Pacific Conference on Communication, August 1993, Taejon, Korea, pp. 6.E.4.1-6.E.4.5.
Go back to Menu
Last Update: August 26 2005.
Any comment for this web is appreciated.
Webmaster