ers can access all these classifiers.
Numerous companies have used such
“deep learning” techniques in commercial products. To classify or separate products based on acceptable or unacceptable
defects, ViDi green software from ViDi Systems (Villaz-St-Pierre, Switzerland; www.vidi-systems.com) allows developers to assign and
label images into different classes after which
untrained images can be classified. In a bottle
sorting application demonstration, Datalogic
(Bologna, Italy; www.datalogic.com) recently
demonstrated how a k-d tree classifier could be
used to identify and sort bottles after test bottles are first presented to the system and key
points in the image automatically extracted
(see “Image classification software goes on
show at Automate,” Vision Systems Design,
May 2015; http://bit.ly/VSD-1505-1).
Developers using CVB Manto from Stem-
mer Imaging (Puchheim, Germany; www.
stemmer-imaging.de) also do not need to
select the relevant features in an image prior
to classification. Using extracted texture,
geometry and color features, captured data is
presented to an SVM for classification. Simi-
larly, the NeuralVision system from Cyth Sys-
tems (San Diego, CA, USA; www.cyth.com)
is designed to allow machine builders with
no previous image processing experience to
add image classification to their systems (see
“Machine learning leverages image classifica-
tion techniques,” Vision Systems Design, Feb-
ruary 2015; http://bit.ly/VSD-1502-1).
By applying multiple image classifiers
on extracted data, developers can deter-
mine whether the extracted features are
good enough to determine the specific fea-
tures of the product being analyzed. If not,
then different types of features may need to
be extracted. Because of this, some compa-
nies offer software packages that allow mul-
tiple classifiers to be developed and tested.
One such toolkit, perClass from PR Sys
Design (Delft, The Netherlands; www.
perclass.com), offers multiple classifiers to
allow developers to work interactively with
data, choose the best features within the data
for image classification, train the numerous
types of classifiers and optimize their perfor-
mance (Figure 4).
Many deep learning resources are now
available on the Web. Two of the most inter-
esting of these are Tombone’s Computer
Vision Blog ( www.computervisionblog.com),
a website dedicated to deep learning, com-
puter vision and AI algorithms and The Jour-
nal of Machine Learning Research (JMLR;
www.jmlr.org), a forum for the publication of
articles on machine learning.
However, while such deep learning
approaches can be used to develop applications such as handwriting recognition,
remote sensing and fruit sorting, they will
always have a limited accuracy, making classifiers less applicable where, for example, parts
need to be measured with high accuracy or
aligned for assembly or processing, or for precision robotic-guidance applications.