Wavelet Multiscale Edge Detection Using An ADALINE Neural Network To Match Up Edge Indicators

Abstract

The detection of sudden changes or discontinuities in data is an important issue in digital image processing. Such changes are often referred to as edge information or just edges. Finding edges are essential to many scientific areas ranging from computer vision to target detection. Not only must the detector be able to find the edges, but it must also be able to detect them in the presence of noise. Many edge-detecting algorithms perform well, but many times these algorithms break down in noisy conditions. One possible solution is to take advantage of the multiscale nature of the wavelet transform to detect edges in noisy conditions. This paper explores one possible method of extracting edge information in two-dimensional sidescan acoustic backscatter imagery using a Wavelet Multiscale Edge Detector (WMED). The WMED uses a wavelet transform to generate coefficients and break down a signal into frequency bands at different levels. Scaling a wavelet, or short waveform, with a scale factor and shifting its position produces these levels. Noise present at low levels is smoothed out and disappears at higher levels. The WMED examines and matches up large magnitude high frequency coefficients, called local maxima, over many different levels to detect edges. To enhance the ability of the detector to operate in very noisy conditions, the WMED is modified to use an ADALINE (ADAptive LInear NEuron) neural network that adapts to match up edge indicators across multiple wavelet levels. The ADALINE uses the least mean squared (LMS) learning rule to minimize the mean square error. The LMS algorithm is able to optimize the decision boundaries of the network. This makes the boundaries more effective in the presence of noise. This paper will test the capability of the ADALINE to match up the edge indicators in noisy two-dimensional sidescan imagery.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Dec 14, 2001
Accession Number
ADA400838

Entities

People

  • James A. Hammack
  • Marlin L. Gendron

Organizations

  • United States Naval Research Laboratory

Tags

Communities of Interest

  • Materials and Manufacturing Processes

DTIC Thesaurus Topics

  • Algorithms
  • Boundaries
  • Change Detection
  • Coefficients
  • Computer Vision
  • Computers
  • Detection
  • Detectors
  • Digital Image Processing
  • Digital Images
  • Frequency
  • Image Processing
  • Neural Networks
  • Signal Processing
  • Target Detection
  • Two Dimensional
  • Wavelet Transforms

Fields of Study

  • Engineering

Readers

  • Atmospheric Science / Meteorology, specifically Wind Wave Turbulence.
  • Computer Vision.
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Neural Networks