kimi's research

Person Recognition System by using Partial Face Images

KIMINORI SATO My HomePage is here.

Abstract

This paper describes a person recognition system that uses partial face images (for example, eye and ear images) for input data. This system has a database of stored individual weights for each person. The weights are trained by a feed-forward neural network that consists of three layers: input layer, hidden layer, and output layer. The network uses a standard back-propagation(BP) learning algorithm. The network contains two output units. One is a recognition unit for recognizing registered persons, and the other is a rejection unit for rejecting unregistered persons (unfamiliar data). Six partial face images for each person are acquired. The first three images are used to train weights for sample images, and the latter three are used to test the accuracy of recognition. From the experimental results of person by partial face (ear) image data of 20 registered persons, a high recognition rate of 93% was obtained, and an error rate of 0% was obtained for the recognition of unregistered persons by using the ear image data. In the case of 50-100 registered persons, a more than 95% recognition accuracy was obtained.

Keywords
 partial face images, back-propagation learning algorithm, weighting factors, a recognition unit, a rejection unit

Summary

  1. Method using a feed-forward neural network
  2. A standard back-propagation learning algorithm
  3. Three layers: input layer, hidden layer, and output layer
  4. Two output units: a recognition unit and a rejection unit
  5. A database of stored individual weighting factors for each person
  6. Training Stage and Testing Stage of registrant A or Testing Stage of registrant D
  7. Training Stage and Database and Testing Stage
  8. Image acquisition system and how to capture face images
  9. Partial face images (eye and ear images) for input data
  10. Extraction Method by using Sobel operater from this image and Extraction by human
  11. Mosaic preprocessing for input unit of a neural network
  12. We have a database of 130 person face image data.
  13. A more than 95% recognition accuracy for registered persons
  14. A more than 98% rejection rate for unregistered persons (unfamiliar data)

My HomePage is here.