Decision Trees for Computer-Aided Detection and Classification (CAD/CAC) of Mines in Sidescan Sonar Imagery

PDF

Authors
  1. Myers, V.L.
Corporate Authors
Defence R&D Canada - Atlantic, Dartmouth NS (CAN)
Abstract
The intuitive training process and easy interpretability make decision trees in attractive solution for comptuer-aided detection and classification (CAD/CAC) of underwater mines in sidescan sonar imagery. Using real sonar data, performance of the decision tree induction process is evaluated using different segmentation methods, that is separating the image into highlight, shadow and background regions. The effect of towfish pitch correction is also investigated. It is argued that the impact of the chosen segmentation algorithm on classification accuracy is minimal in the case of target/clutter discrimination but becomes more important when carrying out target classification, i.e. choosing between cylinder, manta and sphere classes. The induction process is then extended to account for changing prior class probabilities and misclassification costs. Various changes are made with consequences ranging from high sensitivity to relative insensitivity to costs, and a case is made for the use of ROC (Receiver Operating Characteristic) analysis when comparing CAD/CAC methods. Finally, the progression from detection to classification stages is examined.

Il y a un résumé en français ici.

Keywords
Computer aided detection;Target identification;ATR (Automatic Target Recognition);Computer aided classification;Automatic target detection;ATR evaluation;Decision trees
Report Number
DRDC-ATLANTIC-TM-2002-144 — Technical Memorandum
Date of publication
01 Dec 2002
Number of Pages
53
DSTKIM No
CA022087
CANDIS No
518890
Format(s):
CD ROM

Permanent link

Document 1 of 1

Date modified: