Yann LeCun's Home
Publications
OCR
 
LeNet-5 Demos 

Unusual Patterns 
unusual styles 
weirdos 

Invariance 
translation (anim) 
scale (anim) 
rotation (anim) 
squeezing (anim) 
stroke width (anim) 

Noise Resistance 
noisy 3 and 6 
noisy 2 (anim) 
noisy 4 (anim) 

Multiple Character 
various stills 
dancing 00 (anim) 
dancing 384 (anim) 

Complex cases (anim) 
35 -> 53 
12 -> 4-> 21 
23 -> 32 
30 + noise 
31-51-57-61

Multiple Character Recognition

Convolutional Neural nets with a large enough input to receive multiple objects and with multiple outputs are called Space Displacement Neural Networks (SDNN). The LeNet-5 SDNN can recognize character strings in one swoop without requiring prior segmentation.