Automated Prediction of Popularity

clangedbivalveAI and Robotics

Oct 19, 2013 (3 years and 7 months ago)

74 views

Artificial Neural Network For
Automated Prediction of Popularity
of Digitized Images

David Oranchak

doranchak@gmail.com

Objective


Flickr.com ranks photographs based on their
popularity in the site’s user base.


“Interesting”: An image with high rank


“Not Interesting”: An image with low or
nonexistent rank


For any image, can a neural network predict
which group it belongs to?

Approach


Obtain sample images from Flickr.com


559 total samples in the training set


“Very Interesting”: Ranking in the top 25


“Somewhat Interesting”: Ranking between 300 and
500.


“Not Interesting”: No ranking data assigned by Flickr


“Very Interesting”: 25 samples

“Somewhat Interesting”: 11 samples

“Not Interesting”: 36 samples

Approach


Input data sets based on original image data


Raw pixel data, resampled for performance reasons


10x10 RGB pixels


20x20 Grayscale pixels


Color analysis data


One
-
dimensional color counts (histogram)


RGB: three channels, 256 entries per channel


Gray scale: one channel (luminosity), 256 entries


Texture data


Contrast, correlation (inertia), dissimilarity, energy, entropy,
homogeneity, correlation matrix sum, symmetry


Input data derived using
JIU
, a free set of Java image
tools

Approach


Select a suitable neural network architecture


Feedforward backprop architecture?


Result: difficult to train based on input data


Hard to determine suitable number of hidden neurons


Kohonen unsupervised learning?


Result: outputs do not naturally cluster based on
“interestingness”


No mapping between clusters and desired outputs.


Counter Propagation Network?


Result: Very easy to train on input data.

Approach


Training the CPN


559 input patterns


221 patterns for “Very Interesting”


88 patterns for “Somewhat Interesting”


250 patterns for “Not Interesting”


Network simulated using CPN algorithm in JavaNNS, the
Java
-
based successor to SNNS.


Five networks trained successfully; one for each type of
input


Raw RGB pixel data, raw gray scale pixel data, 1D RGB
histogram, 1D gray scale histogram, texture

Experiment 1: Comparison against
Flickr images with known rankings


2381 images from 67 different days obtained
from Flickr


1373 “Very Interesting” images


557 “Somewhat Interesting” images


451 “Not Interesting” images


Experiment 1: Comparison against
Flickr images with known rankings


Results:


32% error rate when at least one network classifies
images as “Very Interesting”


28% error rate when at least two networks classify images
as “Very Interesting”


23% error rate when at least three networks classify
images as “Very Interesting”


14% error rate when at least four networks classify
images as “Very Interesting”


3% error rate when all five networks classify images as
“Very Interesting”

Experiment 1: Comparison against
Flickr images with known rankings


Results are greatly improved when we combine the
categories “Very Interesting” and “Somewhat
Interesting” into a single category: “Interesting”


When one network classifies: 9% error rate


When two networks classify: 9% error rate


When three networks classify: 7% error rate


When four networks classify: 4% error rate


When five networks classify: 2% error rate


Downside: As number of networks go up to reduce
noise, number of missed “Interesting” photos goes
up.

Experiment 2: Flickr photos with
unknown rankings


250 photos sampled at random from recently
uploaded Flickr photos


All five networks classify “Interesting” for 14
of the 250 photos


Experiment 2: Flickr photos with
unknown rankings


Result

Experiment 2: Flickr photos with
unknown rankings





Relaxing the constraint to 4 out of 5 networks
produces 57 images

Experiment 2: Flickr photos with
unknown rankings

Experiment 2: Flickr photos with
unknown rankings


Very subjective results. In my opinion, most
of the photos are interesting!


Experiment 3: Personal photo
collection


2912 samples from personal photo collection


When all 5 networks classify “Interesting”, 98
images result.


Flickr results are better. Personal collection
experiment resulted in many “ordinary
-
looking”
photos.


Test data setup may contribute to lack of success in
this case (resizing of input photos, differences
between Flickr image management and personal
photo formats)

Conclusions


Current CPN technique is very successful
within the Flick image data at locating
interesting photographs


Further experimentation must be performed to
improve success in locating interesting
photographs outside of Flick


More experimentation and refinement must be
done to improve detection rates and reduce
false positives