HANDWRITING RECOGNITION

1
HANDWRITING RECOGNITION
2
Why Handwriting Recognition ?

• Renewed interest
• Advances in screen/tablet technology
• Electronic Ink - the dynamic processing of
  handwriting
• Office automation
• Usability
• Applications appropriate for handwriting
  recognition

3
PDA Constraints

• CPU
• speed (MIPS rating)
• silicon area (gatecount)
• bandwidth on system bus
• Memory requirement (RAM)
• Secondary storage requirement (disk/ flash EPROM)
• Display resolution, frequency, colour depth
• Power ( uW-mW)

4
Handwriting Properties

• Writing styles
• in increasing order of recognition complexity

5
System Partitioning

• Preprocessing Frontend tasks
• Smoothing, Normalisation, Resampling,
  Segmentation
• Feature extraction, Quantisation/Distance
  Calculation
• Recognition
• HMM, NN, Dynamic Programming, Fuzzy architectures
  etc.
• Postprocessing (optional)
• Context-sensitive lexicon and grammar checks

6
Architectures for Pattern Recognition
7
Optimised Architectures
8
Preprocessing steps Smoothing

• Smoothing / Noise reduction
• Heuristic-based thresholds to exclude anomalies
• Incomplete stroke connection
• Wild Point correction
• Dehooking local inaccuracies near pen-up/down
• Filtering for smoothing

9
Preprocessing steps Normalisation

• Normalisation
• Baseline drift correction
• Deskewing, centre-of-mass location
• Size and Stroke length normalisation
• Resampling (time/space domain)

10
Operations for Pre-processing

• MAC-based operations
• differentiation/integration, filtering,
  resampling, scaling, centre-of-mass location
• Elementary functions
• sqrt, sin, cos, tan, sinh, cosh, exp, ln.
• Elementary transformations
• rotation, shear transformation
• Compound operations
• curvature calculation, Hough transform
• Comparison-based operations
• thresholding, extrema finding, intersection
  checking
• Map to general fixed-point 16-bit architectures
  with varying levels of efficiency !

11
Multiplier/ CORDIC architecture
12
Multiplier/CORDIC Performance

• Algorithm performance
• 22 nseconds
• Synthesis results
• only 11 ?m2
• Benchmarking with UNIPEN
• public-domain datasets digits, characters,
  words
• - 88 Accuracy

13
Summary