Send the link below via email or IMCopy
Present to your audienceStart remote presentation
- Invited audience members will follow you as you navigate and present
- People invited to a presentation do not need a Prezi account
- This link expires 10 minutes after you close the presentation
- A maximum of 30 users can follow your presentation
- Learn more about this feature in our knowledge base article
Transcript of Butterfly Identification
which can be incorporated in a butterfly identification phone app or used on its own.
Develop an automatic classifier which identifies correctly the name of a butterfly species based on a segmented image of the butterfly taken in an uncontrolled environment.
Automated Identification of Butterfly Images
Results and Conclusion
The most significant parts of the system are the feature selection and representation and the classifiers.
The bag-of-words method was used for feature representation and selection. The features investigated are :
combinations of those
The classifiers compared are:
One-against-all Support Vector Machine
4 steps required to classify an image:
1. Feature extraction and representation
2. Codebook construction
3. Bag-of-words representation of an image
4. Training of the classifier
Similar approach is employed as with the colour feature. However, the visual words and their descriptors are different.
More sophisticated classifiers,
used to achieve the optimal results.
Idea: represent images using a fixed set of visual words (vocabulary, codebook); spatially invariant representation.
Feature extraction and representation
A codebook, or a vocabulary, is constructed by clustering all pixels from the dataset into a fixed number of clusters.
Codebook – set of cluster centroids (visual words)
Each pixel in an image is then represented as its closest centroid (visual word) . The whole image is represented as a frequency histogram of the number of codewords occurring in it normalized over #pixels in the image, i.e. an m-dimensional feature vector, where m is the size of the vocabulary.
Originally images are matrices of pixels, with each pixel laying in a 3D space - either the RGB (Red-Green-Blue) or HSV (Hue-Saturation-Value) colour spaces.
As a result, images are represented as vectors using their colour feature and the bag-of-words approach.
Colour-based classification: unreliable, sensitive to illumination changes, not discriminative enough.
Pixels, described with RGB values
Circular patches, represented with SIFT descriptors
. . .
. . .
colourSIFT - Extract descriptors from each channel and concatenate them together
histogram concatenation - concatenate colour and shape histograms
simpler classifier, no training phase, # neighbours optimized with cross-validation.
multi-class SVM -
Illumination, low quality images, difficult to gather data etc.
Conclusion and application
Results are averaged over multiple runs of 10-fold cross validation.
For each feature - size of vocabulary and other classification parameters are tuned in. Overall and per-class performance are observed.
Colour vs shape feature
The shape/pattern feature (above)
outperforms the colour feature (below) on 9/10 classes.
colourSIFT improves accuracy of SVM & KNN;
concatenated histograms - only SVM
Best method: concatenated histograms of colour+colourSIFT
KNN - k and distance measure tuned-in
KNN>SVM with colour feature
13-class dataset (1012 images)
Results on the extended dataset
Although the performance generally deteriorates, combining features makes the accuracy more stable.
Possible development directions:
Collection of scientific data
SVM>KNN with shape/combined
SVM error<<KNN's error
Several projects on Butterfly, Bee or Moth identification achieved comparable or better results, but in a much more controlled environment - supervision, specialist equipment, limited classes, family level classification.
The most notable project is a Flower Identification system:
10 species (830 images)