Signal Processing - July 2016 - 33

approach. On the other hand, pixel-based apps, such as wound
assessment (see the "Chronic Wound Assessment" section) or
morphological image assessment, as in skin cancer detection
algorithms (see the "Melanoma Detection Using Mobile Imaging" section), require a low error rate in the segmentation step
because further processes are applied only to the segmented
regions, or assessments heavily rely on the shape of the segmented region, as in the wound assessment and skin cancer
detection apps, respectively. Therefore, these apps require
an investment in more accurate and, in turn, more expensive
segmentation algorithms.

Feature selection
Many of the advanced assisted-living apps, including most of
the health-monitoring apps, include some level of machinelearning techniques. One of the considerations in designing an
efficient machine-learning approach is feature selection, which
can be performed offline to identify a small set of discriminative features to include in the smartphone-based apps. Feature selection is particularly important for smartphone-based
analysis apps. Because of the computational and memory
constraints of smartphones, it is imperative to identify a small
and computationally inexpensive set of discriminative features
to use for statistical model training and apps. Feature selection can reduce the time and memory requirements for feature extraction. In addition, feature-set reduction can decrease
the model app time as well as the complexity of the classifier,
resulting in lower overall memory and energy consumption of
the app. In the case of health-monitoring apps, using a small
set of features may also improve the classification performance
because this reduces the chance of overfitting. For example,
for skin cancer apps, a large number of features (regardless
of computational and memory costs) may ideally encompass
more information and help with correct diagnosis of a skin
legion. However, in reality, because the number of samples to
train the required statistical model is usually low, the trained
model will easily overfit unless a proper feature selection step
is employed.
It is recognized that combinations of individually good features do not necessarily lead to a good classification performance. An exhaustive approach to find the best feature subset
with, at most, k features from a set of M features (k # M )
would require the examination of a significantly large number
of feature subsets

/ ki - 1 c Mi m.
In many cases, it is prohibitive to exhaustively search for the
optimal feature subset.
A well-known practical feature selection procedure is the
normalized mutual information feature selection (NMIFS)
[10]. The NMIFS is a greedy algorithm that determines the
optimal feature subset, starting with the feature that maximizes relevance with target class C. Given a set of selected
features S m - 1, the next feature fm is chosen such that it maximizes the relevance of fm to target class C and minimizes

FIGURE 1. The skin lesion segmentation using an MST-based method [9].
Incorrect and separate segmented components may be due to image
noise. (Photo courtesy of the National Skin Centre, Singapore.)

the redundancy between it and previous selected features in
S m - 1 . In other words, fm is selected such that it maximizes
G function
G ( fm) = I (C, fm) -

1
| Sm - 1 |

NI ( fm, fs) ,

(1)

fs ! S m - 1

where I is the mutual information (MI) function measuring
the relevance between two variables and is defined as
I (X, Y) =

/ / p (x, y) log
y

p (x, y)
,
p (x) p (y)

and NI is the normalized MI function and is defined as
NI (X, Y) =

I (X, Y)
,
min {H (X), H (Y)}

where H is entropy function.
The additional constraints, such as required memory or
computational cost, of each feature should be added to the
optimization step to comply with the targeted mobile device constraints.
Overall, although feature selection is an expensive and
time-consuming process, it can be performed offline on a
powerful computer, and the final feature set is then used on the
mobile device. Investing in a well-analyzed feature selection
can significantly improve the speed, energy consumption, and
even accuracy of the final app.

Color analysis
One of the challenges in image analysis is to find the appropriate color space for the problem being addressed. Given
a proper color space, a segmentation or clustering algorithm may be used to select the ROI, followed by feature
selection. Most of the systems developed for mobile image
processing in the assisted-living field utilize color content

IEEE SIgnal ProcESSIng MagazInE

July 2016

Table of Contents for the Digital Edition of Signal Processing - July 2016