Sie sind auf Seite 1von 11

Extended Back Propagation Position, Rotation and Scale Invariant Neural Processing (Applications in Vision)

A Need for Invariance

Training patterns

Processing is not invariant (true for most networks)


Neural Network

Extending Back Propagation (Difficulties with PSRI)


Existing Proof Bruce, 1988 (facial recognition experiments)

PSRI - some Approaches


Multiple instance training
Exponential growth in number of exemplars

Exponential growth in network size Exponential growth in training time No generalization

PSRI - some Approaches

Higher Order Networks (HONN)


N2C0 possibilities Sampled, Local, Regional connectivity

Neural Network

PSRI - some Approaches


Hough transform FFT Neo-cognitron Etc.
Pre-processing layer FFT, IMAP, Hough, Neo-cognitron

Neural Network

PSRI - some Approaches


Fast Fourier Transform

FFT
Shift affects phase Rotation affects magnitude Scale (linear shift along ln(Fr) axis

Neural Network

Extending Back Propagation (A Plausible Approach)


y Object ln r

Theta

Mapping from the xy to the ln(r)-theta plane results in a size invariance pattern.

Extending Back Propagation


(A Plausible Approach -- Achieving Size Invariance)
3 2 y -1 0 1 x 2 3 1 0 -1 y 4 3 2 1 0 -1 -1 0 1 x Transformation to the ln r-theta plane 0.4 0.3 ln r ln r 0.2 0.1 0.0 -4 -2 0 Theta 2 4 0.8 0.7 0.6 0.5 0.4 -4 -2 0 Theta 2 4 2 3 4

Extending Back Propagation


(A Plausible Approach -- Achieving Rotation Invariance)
2 1 0 -1 -2 -2 -1 0 x 1 2
ln r y

0.4 0.3 0.2 0.1 0.0 -4 -2 0 Theta 2 4

Rotation of the ln r-theta plane produces an identical rotation in the xy plane

Extending Back Propagation (The Mechanism)


Object Scaled object
...

Four views generated by rotating the resulting mapping

Mapping to the ln r-theta plane

Fully connected back propagation network Rotate ln r-theta plane Error Comparison logic Unacceptable error Shape

Das könnte Ihnen auch gefallen