Saimaa ringed seal fur pattern extraction for identification purposes
Nepovinnykh, Ekaterina (2017)
Julkaisun pysyvä osoite on
http://urn.fi/URN:NBN:fi-fe201705236829
http://urn.fi/URN:NBN:fi-fe201705236829
Tiivistelmä
The Saimaa ringed seal is considered to be endangered and is facing a very high risk
of extinction. he conservation efforts largely depend on the ability to track and monitor
each individual seal. Photo-identification using camera traps has been successfully
used for wildlife monitoring. Each seal has a unique fur pattern that a human expert can
match to a specific seal labeled earlier. This thesis focuses on automatic identification of
Saimaa ringed seals based on fur pattern extraction. This consists of segmentation of an
image with the goal of extracting the seal, extraction of fur pattern from the segmented
seal image and searching for the same seal in the seal database. Two methods of Saimaa
ringed seal identification based on transfer learning are proposed in this work. The first
method involves re-training of the existing convolutional neural network (CNN). The second
method involves using the existing CNN trained for image classification as a means to
extract features from seal images which are then used to train a Support Vector Machine
(SVM) classifier. Both methods are implemented, tested and compared. Both approaches
show good results with total accuracy of 91.2% for CNN and 90.5% for SVM.
of extinction. he conservation efforts largely depend on the ability to track and monitor
each individual seal. Photo-identification using camera traps has been successfully
used for wildlife monitoring. Each seal has a unique fur pattern that a human expert can
match to a specific seal labeled earlier. This thesis focuses on automatic identification of
Saimaa ringed seals based on fur pattern extraction. This consists of segmentation of an
image with the goal of extracting the seal, extraction of fur pattern from the segmented
seal image and searching for the same seal in the seal database. Two methods of Saimaa
ringed seal identification based on transfer learning are proposed in this work. The first
method involves re-training of the existing convolutional neural network (CNN). The second
method involves using the existing CNN trained for image classification as a means to
extract features from seal images which are then used to train a Support Vector Machine
(SVM) classifier. Both methods are implemented, tested and compared. Both approaches
show good results with total accuracy of 91.2% for CNN and 90.5% for SVM.