1. The Assistive System Progress Report 2 Shifali Kumar Bishwo Gurung
James Chou

2. New Additions Use of local and global features
Datastore of image categories (Caltech 101 Categories)
Training classifiers (SVM, random forest)
Text-to-speech synthesizer

3. Histogram of Oriented Gradients
Uses global features to describe the image.
Typically used for human detection.
Used with SVM for image classification.

4. Histogram of Oriented Gradients (Example)

5. Code Structure

6. Support Vector Machine
A probabilistic binary linear classifier which is a part of machine learning
It builds a learning model that assigns new examples into certain categories
It makes classification decision based on the value of linear combinations of the characteristics; making features presented to the machine into feature vectors (ex. pixels)

7. The Classifier ASSEMBLE - Create image data set of three categories
TRIM - Make the number of images equal
PARTITION - Divide the data set into training and validation sets
TRAIN - Create a model for future sets
TEST - Predict the possible category where the new image belongs

8. Sample Categories

10. BagofFeatures Model

11. Classifier Data Average Accuracy
Category
Starfish
Rhino
Cup
Count 86 52 57
Average Accuracy
Training Set: 90% (works with collected features from old images)
Validation Set: 70% (uses old data to use for new data)

12. Results
Training Set Validation Set

13. Code Snippets imgSets = partition(imgSets, minSetCount, 'randomize');
[trainingSets, validationSets] = partition(imgSets, 0.3, 'randomize');
% input an image to be classified into the proper category
img = imread(fullfile(rootFolder, 'unknown', 'image_0001.jpg'));
[labelIdx, scores] = predict(categoryClassifier, img);
% Display the string label categoryClassifier.Labels(labelIdx)

14. HOG with SVM

15. Sample Results
ANS: RHINO ANS: STARFISH ANS: STARFISH

16. Random Forest Functions as an object classifier
A technique of machine learning using an ensemble of decision trees
More stable by averaging trees that grow irregularly to give less variance
Has improved predictive performance by providing less random errors
Utilizes the random subspace method

17. Random Forest

18. Conclusion
Observed different methods of extracting features from images.
The running time with BagofFeatures model took longer than expected.
Hopes of random forest outperforming in speed and accuracy.
Speech-to-text feature is next.

19. Any questions and concerns are welcome.
Thank you!
Any questions and concerns are welcome.