Transcript Image noise filtering using artificial neural network

Image noise filtering using
artificial neural network
Final project by Arie Ohana
Image noise
High frequency random perturbation in pixels
In audio, noise can be a background hiss
Total elimination of noise can rarely be found
Can use blurring for reduction
Many kinds: Additive, Salt & pepper, etc…
Salt & pepper noise
A clean image
S&P noise, Density = 0.1
Artificial Neural Network
A computing paradigm that is loosely
modeled after cortical structures of the brain.
Consists of interconnected processing
elements called neurons.
Achieves its goal by a learning process.
The network will adjust itself, by correcting
the current weights on every input, according
to a predefined formula.
Depends heavily on the expressiveness of
exemplars.
Neural Network / Structure
Output Values
synapse
axon
nucleus
cell body
dendrites
Input Signals (External Stimuli)
A neuron in the brain
Basic perceptron
Multi layers ANNs
Approach and Method
Running exemplars for 50,000 epochs.
Using 4 expressive images
Using 1 hidden layer, with 50 neurons
Input is a given pixel value along with its
surrounding 8 neighbors.
Output is single grayscale value (the
correction).
The Training Set
Complex
gradients
A dichotomy image
Gradients and details
A detailed image
Filtering images / Results
Complex images, comparing to existing methods
Filtering images / Results
Complex images, comparing to existing methods
Filtering images / Results
Complex images, comparing to existing methods
Filtering images / Results
Less complex, more dichotomy images
How about filtering noise from (beautiful) faces?
Artificial simple images
Analysis
It seems that the network used blurring
and whitening (brightening).
When zooming in, we can clearly observe the blurring effect
The brighten method can
clearly be seen
Analysis
Filtering a complex image
The histogram of a typical image.
Grayscale histogram of the image
as produced by the NN. The
damage is pretty large.
Analysis
Filtering a simple image
The histogram of a
dichotomy image.
The histogram the NN
produced which very similar to
the source.
Conclusions
The network used mostly blurring and
brightening
When comparing to existing methods, they
seem preferable
Bear in mind: test cases were mostly very
complex and difficult
Filtering simple dichotomy images was
easy for the network
Future work / Improvements
Problem: noise is being filtered even in
pixels that weren't noised.
Image is heavily corrupted, even with
existing methods for noise reduction.
Solution: build an ANN for recognizing
noise only (should be easy and with small
False alarm).
Use an ANN or other method for filtering noise
locally only.
Future work / Improvements
Find noised pixels
Filter only noised pixels
Noise / No Noise
Greyscale values
Output Values
Input Signals (External Stimuli)
A clean pixel is transparent
Noised image
Filtered image
Questions…