Showing papers on "Convolutional neural network published in 1993"

Globally Trained Handwritten Word Recognizer using Spatial Representation, Convolutional Neural Networks, and Hidden Markov Models

Abstract: We introduce a new approach for on-line recognition of handwritten words written in unconstrained mixed style. The preprocessor performs a word-level normalization by fitting a model of the word structure using the EM algorithm. Words are then coded into low resolution "annotated images" where each pixel contains information about trajectory direction and curvature. The recognizer is a convolution network which can be spatially replicated. From the network output, a hidden Markov model produces word scores. The entire system is globally trained to minimize word-level errors.

Computer-assisted diagnosis of lung nodule detection using artificial convoultion neural network

Abstract: Several fuzzy assignment methods for the output association with convolution neural network are proposed for general medical image pattern recognition. A non-conventional method of using rotation and shift invariance is also proposed to enhance the neural net performance. These methods in conjunction with the convolution neural network technique are generally applicable to the recognition of medical disease patterns in gray scale imaging. The structure of the artificial neural network is a simplified network structure of neocognitron. Two- dimensional local connection as a group is the fundamental architecture for the signal propagation in the convolution (vision type) neural network. Weighting coefficients of convolution kernels are formed by neural network through backpropagated training for this artificial neural net. In addition, radiologists' reading procedure was modeled in order to instruct the artificial neural network to recognize the pre-defined image patterns and those of interest to experts. We have tested this method for lung nodule detection. The performance studies have shown the potential use of this technique in a clinical environment. Our computer program uses a sphere profile double-matching technique for initial nodule search. We set searching parameters in a highly sensitive level to identify all potential disease areas. The artificial convolution neural network acts as a final detection classifier to determine if a disease pattern is shown on the suspected image area. The total processing time for the automatic detection of lung nodules using both pre-scan and convolution neural network evaluation is about 10 seconds in a DEC Alpha workstation.

Postal Address Block Location Using a Convolutional Locator Network

Abstract: This paper describes the use of a convolutional neural network to perform address block location on machine-printed mail pieces. Locating the address block is a difficult object recognition problem because there is often a large amount of extraneous printing on a mail piece and because address blocks vary dramatically in size and shape. We used a convolutional locator network with four outputs, each trained to find a different corner of the address block. A simple set of rules was used to generate ABL candidates from the network output. The system performs very well: when allowed five guesses, the network will tightly bound the address delivery information in 98.2% of the cases.

Application of artificial neural networks for reduction of false-positive detections in digital chest radiographs.

Abstract: A methodology based on the fuzzy set theory and the convolution neural network (CNN) architecture is proposed to tackle the problem of reducing false-positive rate in automatic lung nodule detection. The CNN which simulates human visual mechanism was trained by a supervised back-propagation algorithm based on fuzzy membership functions. The training and testing database consists of image blocks (each 32 x 32 pixels) of suspected lung nodule areas (nodule candidates) which were generated from our pre-scanning program [1]. A linguistic label was assigned to each nodule candidate of the training set, then the label was converted to a membership value through a pre-defined membership function and used as teaching signal (desired outputs) during the network learning. Before the nodule candidate was fed to the network input, it was pre-processed to reduce the complex background noise and the contrast discrepancy resulted from film development. During the network testing phase, a defuzzification process was applied to decipher the trained network's output triggered by the nodule candidate in the testing set. Finally, a Receiver Operating Characteristic (ROC) analysis was used to evaluate the CNN's performance based on the defuzzified output of the testing database. Preliminary results showed an average Az (the performance index) of 0.84 which is equivalent to 0.80 true-positive detection (sensitivity) with an average 2-3 false-positive detections per chest image.

On-line recognition of limited-vocabulary Chinese character using multiple convolutional neural networks

Abstract: The authors present a new feature extraction method together with neural network recognition for online Chinese characters. A Chinese character can be represented by a three-dimensional 12 /spl times/ 12 /spl times/ 4 array of numbers. Multiple conventional neural networks are used for online small vocabulary Chinese character recognition based on this feature extraction method. One hundred character classes were chosen as an example for recognition. Simulation results show that 98.8% and 94.2% of training examples and test examples were correctly recognized respectively. >

Extended backpropagation for invariant pattern recognition neural networks

Abstract: This paper presents some improvements on a neural network structure composed by a multilayer perceptron (MLP) with a preprocessing neural net, in order to perform translation, rotation and scale invariant pattern recognition. The preprocessing network has been modified and backpropagation (BP) has been generalized for training the preprocessing net as well as the multilayer perceptron. The new structure and weight selection procedure lead to a more noise tolerant neural net which also performs a better pattern classification.

An initialization method for multilayer neural networks as function approximators

Abstract: A multilayer neural network and its initialization method, which takes the distribution of given training patterns into consideration, are proposed. The network is composed of four layers, and the role of each layer in the network is specialized by analyzing the internal information processing of the network. The size of the network, initial values of weights, and parameters defining the characteristics of the nonlinearities of processing units in hidden layer are determined from a selected portion of the given training patterns. With these initial conditions, the performance of the network is further improved by the general error backpropagation learning process. The proposed model and method give any desired mapping performance with smaller network size and faster learning speed than obtained with conventional multilayer neural networks and a random initialization technique. To show the usefulness of the proposed network, an illustrative mapping is realized by the proposed algorithm.