Development of an automated image analysis system to detect beluga whales in aerial photographs
One aspect of monitoring the population of beluga whales, and other marine mammal species, is counting a sample of the population from aerial photographs (or negatives). Using image processing and pattern recognition techniques, a software system for detecting and classifying beluga whales in digiti...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
Memorial University of Newfoundland
2006
|
| Subjects: | |
| Online Access: | https://research.library.mun.ca/11048/ https://research.library.mun.ca/11048/1/Mills_Jason.pdf |
| Summary: | One aspect of monitoring the population of beluga whales, and other marine mammal species, is counting a sample of the population from aerial photographs (or negatives). Using image processing and pattern recognition techniques, a software system for detecting and classifying beluga whales in digitized aerial photographs and negatives is developed. The image processing component includes algorithms to create a mask to cover "unreadable" areas (e.g. land and sun glare), segment whales, and generate feature data for segmented objects. The segmented objects are classified and presented to the user in an interactive GUI (graphical user interface) for final conformation and quality control. -- A fundamental step in developing a good pattern recognition system is to choose and optimize a classifier. To this end, the support vector machine (SVM) classifier is compared against a traditional quadratic discriminate classifier. To optimize the classifiers, a genetic algorithm (GA) for feature selection and classifier parameter calibration is used. An obstacle in applying GAs to any problem is selecting values for the fundamental GA control parameters. This is addressed using design of experiments (DOE) to systematically analyze the GA and derive a statistical model from which the parameters can be calculated. It is demonstrated that GAs are a good method to optimize SVMs via feature subset selection and SVM parameter calibration. |
|---|