Neural Network Architectures for General Image Recognition.
Abstract
As part of Lincoln Laboratory's research on neural network technology, a general-purpose machine vision system was designed that can learn to recognize diverse objects. This system models human vision, primarily with neural networks, and the learning is by example. The system was tested on two disparate classes of objects, military vehicles and human cells, with video images of natural scenes. These objects were chosen because large databases are available and because most researchers judge them to be unrelated. The system was trained and tested with 40 images of military vehicles. After training with 18 images, the system was able to recognize the tanks, howitzers, and armored personnel carriers of the remaining images without error. Pathologists at Lahey Clinic Medical Center collaborated in the cytology study where the system was trained and tested on 156 cell images from human cervical Pap smears. After training with 118 images, the system correctly classified all of the other cells as normal or abnormal (that is, precancer). These results axe preliminary because the number of military vehicles and Pap smear samples was small. Nonetheless, the results are extremely encouraging and clearly indicate that additional development of the system is warranted. We note that the architecture of the system is applicable to many civilian and military tasks. The application of the system to a specific task requires appropriate training.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 21, 1992
- Accession Number
- ADA256017
Entities
People
- Karl G. Heinemann
- Paul N. Dicaprio
- Robert L. Harvey
Organizations
- Massachusetts Institute of Technology