Semi-automatic ground truth generation using unsupervised clustering and limited manual labeling: Application to handwritten character recognition
Download as PPTX, PDF0 likes942 views
The document presents a semi-automatic method for generating ground truth labels using unsupervised clustering combined with limited manual labeling, aimed at improving handwritten character recognition. It describes various feature representation techniques and evaluates the performance of clustering and labeling strategies on datasets like MNIST and Lampung. The results demonstrate competitive classification performance with minimal human intervention, emphasizing the efficiency of the proposed approach.
Semi-automatic ground truth generation using unsupervised clustering and limited manual labeling: Application to handwritten character recognition
- 1. SEMI-AUTOMATIC GROUND TRUTH GENERATION USING UNSUPERVISED CLUSTERING AND LIMITED MANUAL LABELING: APPLICATION TO HANDWRITTEN CHARACTER RECOGNITION SzilárdVajda,Yves Rangoni, Hubert Cecotti Pattern Recognition Letters, 2015 1
- 2. Ground-truth generation UnlabeledLabeled • Usually, real world data is not labeled. • Large data collections need accurate labels. 2
- 3. Labeling strategy Unsupervised clustering Labeling by human expert the closest real data point for each centroid is labeled 3 Image dataset Pixels Profiles LBP Randon Encoder Feature representations
- 4. Image dataset Pixels Profiles LBP Randon Encoder Feature representations Label a Label c Label b Labeling strategy each point inherits a label of its cluster 4
- 5. Input image Pixels Profiles LBP Randon Encoder Feature representations 5 8 5 5 5 5 Final label Consensus voting / Majority voting Labeling strategy 5
- 6. Feature representations • Raw pixels – Pixel intensity in raw images • Profiles (upper/lower/left/right) – only considers outer shape of the character – i.e. consider the distance between the upper horizontal line and the closest pixel to the upper boundary of the image • Local Binary Patterns (LBP) – local texture and rotation invariant representation 6 L. Heutte, T. Paquet, J.V. Moreau, Y. Lecourtier, C. Olivier, A structural/statistical featurebased vector for handwritten character recognition, Pattern Recognit. Lett. 19 (7) (1998) 629–641.
- 7. Feature representations • Randon transform – takes multiple and parallel-beam projections of the image from different angles • Encoder network – a special kind of deep learning architectures – data-driven 7
- 8. • Definitions • Voting scheme: consensus, majority voting Classifiers the number of patterns that should be assigned to the i-th class the number of patterns that are assigned to the class after classification 𝑁𝑝 = 𝑁𝑑𝑒𝑐 + 𝑁𝑟𝑒𝑗 𝑁𝑑𝑒𝑐: patterns that have a class assigned, 𝑁𝑟𝑒𝑗: patterns that have no assigned patterns 𝑁+ / 𝑁−: patterns that have been correctly / incorrectly classified 8
- 9. Classifiers • Unsupervised clustering – K-means clustering (Lloyd algorithm) – Self Organizing Map (SOM) : a special type of neural network trained in an unsupervised fashion, to produce a two-dimensional mapping of the input data – The Growing Neural Gas (GNG) : no constraints on the topology contrary to the SOM • Supervised classification – The k-nearest neighbor (k-nn) classifier 9
- 10. Classifiers • Evaluation • measures combine inter-class and intra-class variances • measures the reliability of the labeling strategy X: total numbers of vectors to be clustered 1 0
- 11. Dataset • MNIST – Arabic digits – 10 classes (0,1,…,9) – 60,000 training / 10,000 test images • Lampung – multi-writer handwritten collection produced by 82 high school students from Bandar Lampung, Indonesia – 20 character classes – 23, 447 characters for training – 7,853 characters for the test 1 1
- 12. Results • Performance of features 1 2
- 13. Results • Compactness of clustering techniques 1 3
- 15. Results • Labeling performance – Majority / consensus voting: at least 3 methods / 5 methods provide the same label 1 5
- 16. Results • Labeling performance 1 6 Competitive performance is shown with few human-labeled samples
- 17. Results • Classification performance – against several Monte Carlo simulations (100 times) which pick random samples from complete training set. 1 7
- 18. Results • Classification performance (different voting) – A fully connected multi layer perceptron classification 1 8 96.69 96.74 96.77 The network is more sensitive to the samples with wrong labels
- 19. Conclusion • Semi-automatic labeling scheme with minimal human involvement. • The newly discovered labels with this labeling scheme are compared in a k-nn scheme, with randomly selected samples and the complete data (all labeled). 1 9
- 20. Thank you ! Q & A 2 0