Showing papers on "Convolutional neural network published in 1994"

A Convolutional Neural Network Hand Tracker

Abstract: We describe a system that can track a hand in a sequence of video frames and recognize hand gestures in a user-independent manner. The system locates the hand in each video frame and determines if the hand is open or closed. The tracking system is able to track the hand to within ±10 pixels of its correct location in 99.7% of the frames from a test set containing video sequences from 18 different individuals captured in 18 different room environments. The gesture recognition network correctly determines if the hand being tracked is open or closed in 99.1% of the frames in this test set. The system has been designed to operate in real time with existing hardware.

Word-level training of a handwritten word recognizer based on convolutional neural networks

Abstract: We introduce a new approach for online recognition of handwritten words written in unconstrained mixed style. Words are represented by low resolution "annotated images" where each pixel contains information about trajectory direction and curvature. The recognizer is a convolutional 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.

Axially symmetric neural network architecture for rotation-invariant pattern recognition

Abstract: We propose an axially symmetric neural network architecture capable of detecting orientations for rotation-invariant pattern recognition. This is a class of multilayer perceptrons, where the synaptic weights "in parallel" between the lower layer and the successive upper layer all have the same values (i.e., symmetric with respect to the principal axis of the network architecture). This network can be trained by the backpropagation procedure under the weight constraints using any patterns (e.g. 10-digit or 26 alphabet characters) in a standard position, which automatically makes it possible to recognize the test patterns with different orientations (i.e., rotated patterns), simultaneously detect the orientation. >

Design of digital accelerators for backpropagation

Abstract: This paper proposes an efficient technique for implementing artificial neural networks (ANNs). This technique is utilized to design two fast neurocomputers; FMAT1 and FMAT2. The paper concentrates on the recall and learning phases of multilayer perceptrons with backpropagation learning. FMAT1 requires less hardware but is appropriate for the recall phase only. With a small additional cost, FMAT2 adds the capability of learning. When compared to other techniques in the literature, FMAT1 and FMAT2 exhibit superior performance. They provide a better connections per unit time measure. To compute a neural network having N Neurons in its largest layer, These two architectures require O(log N) processing time. Another major virtue of these architectures is their ability to pipeline multiple patterns which further improves performance.