Information Technology Reference
In-Depth Information
Fig. 11.2
Feature extraction steps
Neural Network Simulator
,Zelletal.
2003
) and standard back-propagation. The
results presented in this chapter concern artificial neural networks with one input
unit per feature, 12 units in the hidden layer, and 2 units in the output layer (one
for each category). A training pattern specifying an output of (1; 0) indicates that
the corresponding image belongs to the first set. Likewise, a training pattern with
an output of (0; 1) indicates that the corresponding image belongs to the second set.
The parameters for the classifier and FE were established empirically in previous
experiments.
The experiments presented in this section concern classification tasks of different
nature: aesthetic value prediction, author identification and popularity prediction.
All the results presented in this section were obtained by the same AJS, trained in
different ways. Finally, we describe the integration of this AJS with an evolutionary
image generation system.
11.3.2.1 Feature Extraction
In this section we describe the feature extraction process.
The feature extraction can be summarised to the following steps (see Fig.
11.2
):
(i)
Pre-processing,
which includes all the transformation and normalisation opera-
tions applied to a given input image; (ii)
Metrics application
, that is, the application
of certain methods based on statistical measurements and image complexity esti-
mates; (iii)
Feature building
, the extraction of results from the metrics applied in
order to build the image feature set.
Pre-processing
The images from a dataset are individually submitted to a series
of transformations before being analysed. A given input image is loaded and resized
to a standard width and height of 256
256 pixels, transformed into a three-channel
image in the RGB (red, green and blue) colour space, with a depth of 8 bits per
channel and all the pixel values are scaled to the [0
,
255] interval. This step ensures
that all input images share the same format and dimensions.
Next, the image is converted into the HSV (Hue, Saturation and Value) colour
space and its HSV channels are split. Each of these channels is stored as a one-
channel greyscale image. From here on, we will refer to these images as H, S and
V channel images. A new greyscale image is also created by performing a pixel by
pixel multiplication of S and V channels and scaling the result to [0
,
255]. From
now on, we will refer to this image as the CS (Colourfulness) channel image.
The images resulting from these operations are subject to transformation op-
erations. The current version of the FE supports seven transformations: no filter,
×
