Sketch recognition for interactive web apps

Jesús Martínez-Blanco

de.linkedin.com/in/chumo

@jmb_jesus

@chumo

Digit recognition

LeNet 1 (1993):

The first CNN with good speed and accuracy to recognise digits.

using Convolutional Neural Networks

Yann LeCun

(Bell Labs)

Soon used to read the numerical amounts on checks.

Neural Net for Handwritten Digit Recognition

http://myselph.de/neuralNet.html

Trained network from:

For web apps,

Random Forest or SVM

something fast and robust

input

feature engineering

machine learning

output

randomforest.js

svm.js

Andrej Karpathy

@karpathy

sample
normalize
standarize

Coordinates:

RAW

to be insensitive to the drawing speed

to be insensitive to absolute positions

Sketch recognition

feature engineering

[
  {"label":0, "features":[4.29, 4.47, 1.80,..., 2.06]},
  {"label":1, "features":[4.85, 4.73, 3.80,..., 2.36]},
  {"label":2, "features":[4.75, 2.34, 3.12,..., 5.13]},
  {"label":3, "features":[1.40, 0.92, 4.49,..., 5.09]},
  .
  .
  .
  {"label":9, "features":[2.18, 2.04, 5.83,..., 2.96]}
]