Search PubMed⌕ Search

Biomedical subjects

H C Fu

Publications and source records attributed to H C Fu.

14 recordsLinked to original sources
PubMed1999

Recognition of handwritten similar Chinese characters by self-growing probabilistic decision-based neural network.

Recognition of similar (confusion) characters is a difficult problem in optical character recognition (OCR). In this paper, we introduce a neural network solution that is capable of modeling minor differences among similar characters, and is robust to various personal handwriting styles. The Self-growing Probabilistic Decision-based Neural Network (SPDNN) is a probabilistic type neural network, which adopts a hierarchical network structure with nonlinear basis functions and a competitive credit-assignment scheme. Based on the SPDNN model, we have constructed a three-stage recognition system. First, a coarse classifier determines a character to be input to one of the pre-defined subclasses partitioned from a large character set, such as Chinese mixed with alphanumerics. Then a character recognizer determines the input image which best matches the reference character in the subclass. Lastly, the third module is a similar character recognizer, which can further enhance the recognition accuracy among similar or confusing characters. The prototype system has demonstrated a successful application of SPDNN to similar handwritten Chinese recognition for the public database CCL/HCCR1 (5401 characters x200 samples). Regarding performance, experiments on the CCL/HCCR1 database produced 90.12% recognition accuracy with no rejection, and 94.11% accuracy with 6.7% rejection, respectively. This recognition accuracy represents about 4% improvement on the previously announced performance. As to processing speed, processing before recognition (including image preprocessing, segmentation, and feature extraction) requires about one second for an A4 size character image, and recognition consumes approximately 0.27 second per character on a Pentium-100 based personal computer, without use of any hardware accelerator or co-processor.

China↗
PubMed1994

A fuzzy neural network for knowledge learning.

This paper presents a fuzzy neural network for learning the knowledge of a fuzzy logic rule-based system. The network contains five layers: an Input Layer, Membership-function Layer, AND Layer, OR Layer, and Defuzzification Layer. We propose a backpropagation-like learning algorithm to train this neural network to acquire the fuzzy rules and to fine-tune the knowledge on the parameters of AND and OR nodes. Compared with methods other than the gradient descent search, the proposed learning process acquires more precise knowledge. In addition, the functions of the AND and OR nodes in the network are formulated with the minimum or maximum operations, respectively. Therefore, the adjustments of the learnable weights (parameters) can be focused on the dominant terms related to the (minimum/maximum) operations. The convergence time for the proposed learning algorithm is much faster than that for conventional backpropagation algorithms. In summary, the learnable weights (parameters) of the network are adjusted very quickly to obtain precise knowledge. Simulation results show that in learning the truck backer-upper problem, our network completes the training procedure in only several dozen epochs with an error rate of less than 1%.

Algorithms↗
PubMed1994

On the classification capability of a dynamic threshold neural network.

This paper proposes a new type of neural network called the Dynamic Threshold Neural Network (DTNN) which is theoretically and experimentally superior to a conventional sigmoidal multilayer neural network in classification capability. Given a training set containing 4k + 1 patterns in Rn, to successfully learn this training set, the upper bound on the number of free parameters for a DTNN is (k + 1)(n + 2) + 2(k + 1), while the upper bound for a sigmoidal network is 2k(n + 1) + (2k + 1). We also derive a learning algorithm for the DTNN in a similar way to the derivation of the backprop learning algorithm. In simulations on learning the Two-Spirals problems, our DTNN with 30 neurons in one hidden layer takes only 3200 epochs on average to successfully learn the whole training set, while the single-hidden-layer feedforward sigmoidal neural networks have never been reported to successfully learn the given training set even though more hidden neurons are used.

Algorithms↗