10
Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Embed Size (px)

Citation preview

Page 1: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Identification and Enumeration of Waterfowl using Neural

Network Techniques

Michael Cash

ECE 539 Final Project

12/19/03

Page 2: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Background (i)

• Annual waterfowl surveys required for population estimations

• Many techniques used:– Area search (direct count)– Point count– Aerial Survey– Hunter seasonal survey

• Aerial surveying is costly but covers the most ground

Page 3: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Background (ii): Aerial Surveys

• Aerial pilots cruise along selected route, record number and species of birds seen

• Cost prohibits large number of pilots; small sample of actual population recorded

• Subject to human discretion and error

• Birds in flight are difficult to identify and count

Page 4: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Proposal

• Digital images can be processed to count the number of ducks in a flock

• One way to do this is with a k-means algorithm

• Once the locations of ducks in the image are known, a perceptron classifier could identify the species and/or gender

Page 5: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Image Pre-Processing

• Digital image of flock of ducks must reduce effects of background for clustering scheme to work

• Trees, vegetation removed by cropping (leave only sky and/or water)

• Median pixel value of entire image is set as nominal background

• Small range about nominal background set to provide for lighting & shadows

Page 6: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

K-Means Clustering

• 250 randomly placed cluster centers

• Algorithm moves cluster centers until converged (centers no longer move) to <1E-30

• Empty clusters are removed

• Number of clusters remaining is number of ducks

Page 7: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Example Results

Cropped Image Feature Vectors (red)Cluster Centers (yellow)

59 observed birds 52 predicted birds

Page 8: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Example Results

Cropped Image Feature Vectors (red)Cluster Centers (yellow)

16 observed birds 25 predicted birds

Page 9: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Results Summary

• The algorithm produced results within 50%, but error can be reduced by:– Spacing birds (reducing clustered birds)– More pixels– Even background

Page 10: Identification and Enumeration of Waterfowl using Neural Network Techniques Michael Cash ECE 539 Final Project 12/19/03

Conclusion

• Proof of concept: one unaltered algorithm was used for many images, and produced desirable results for counting birds

• Perceptron classifier was not added due to lack of quality in images, but could be added in the future