Procedia Computer

Science
Procedia ComputerScience
Procedia Computer Science001(2010)
(2012) 913–919
1–7
www.elsevier.com/locate/procedia

International Conference on Computational Science, ICCS 2010

Introductory computational science using MATLAB and image

processing
D. Brian Larkins1 , William Harvey
Dept. of Computer Science and Engineering
The Ohio State University
Columbus, OH 43210
{larkins,harveywi}@cse.ohio-state.edu

Abstract
We describe a new course designed to introduce engineering students to computational thinking. One of the most
significant challenges in teaching an introductory-level applied computing course is that students are not expected to
have substantial math and science experience. To address this, we have developed exercises which use edge detec-
tion and basic image processing to motivate the use of programming MATLAB in a non-trivial scientific application.
MATLAB is a popular high-level programming language and environment which supports a wide range of compu-
tational science applications. MATLAB has strong support for operating on image data, which allows us to balance
solving practical engineering problems with basic core concepts of computer science such as functional abstraction,
conditional execution, and iteration.
c 2012 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
⃝

1. Introduction
The use of computers for simulation and data analysis has become widespread throughout many engineering dis-
ciplines. As a result, engineering departments are requiring students to take courses which provide an introduction
to computational thinking and applied computer programming. The need for students to have familiarity with al-
gorithmic thinking, programming concepts and their application to science and engineering problems spans many
disciplines and departments. As a result, an introductory course has been created and offered through the Department
of Computer Science and Engineering at Ohio State. Designing a general, introductory class which effectively de-
livers a mix of basic programming concepts and scientific applications is a challenge. Such a course must strike a
balance between covering the core concepts of computation with the need for practical skills to solve real engineering
problems.
The students in this course come from a variety of science and engineering backgrounds: from electrical, civil, and
systems engineering, as well as materials science and other fields. To allow a wide range of students from a diverse
range of disciplines to take this course early in their programs, students are only required to have had at least one
calculus course, with no other prerequisites. As such, it is difficult to develop exercises and other experiential course
components which reflect actual applications related to a student’s chosen field.

1 corresponding author

1877-0509 ⃝c 2012 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
doi:10.1016/j.procs.2010.04.100
914 D.B. D.
Larkins, W. Harvey / Procedia Computer Science 1 (2012) 913–919
B. Larkins et al. / Procedia Computer Science 00 (2010) 1–7 2

Regardless of the math and science background of a student or their intended major, an introductory course in
computational science must teach practical skills that will be useful to every student. While it is possible to solve
many science and engineering problems using low-level programming languages, it is much more common in practice
to use high-level tools which support problem-solving within a specific domain. Ideally, students should be exposed
to applied problem solving with computational environments that they would see later in their education or vocation.
To address the tension between providing basic computer science knowledge and the need for practical experience
with applied scientific computing, this course is taught using the MATLAB programming language environment.
MATLAB is a powerful tool which has seen widespread adoption throughout industry and academia. Since MATLAB
provides a complete high-level programming language, it can be used as a basis for teaching core CS concepts. In
this paper, we describe the outline of an introductory course suitable for teaching science and engineering students
computational thinking. Our approach is based on two key insights. First, MATLAB is a tool which provides a rich
environment for both exploring general computing concepts and solving domain-specific applied problems. Second,
that some applied image processing problems can be intuitively understood with a basic calculus background and are
similar in spirit to other applied science and engineering problems.
We make the following contributions: First, we describe a course offered at Ohio State which covers both general
computer science concepts and applied scientific computing using MATLAB. Second, we demonstrate how image
processing can be used as a model scientific application. We provide specific exercises based around detecting the
edges in an image which give students experience solving realistic problems, but do not require substantial math
background. Lastly, we demonstrate the effectiveness of these exercises with results from a student survey.

1.1. MATLAB as a Computational Tool

MATLAB is software environment which supports high-level programming, simulation, and visualization of sci-
ence and engineering applications. MATLAB is a useful introductory tool for teaching computational science for
several reasons: MATLAB enjoys popularity in many different science and engineering disciplines in both industry
and academia [1]. It is a multi-platform system, which is scalable from laptops all the way up to high-performance
computing clusters [2]. There is a broad support base, with an active user community, sample code repositories, and
the availability of many textbooks. MATLAB also supports a large number of specialized toolboxes, such as support
for advanced statistics, partial differential equations, signal processing, etc.
A key benefit of using MATLAB is that while much of the programming can be done at a high-level, it also
supports general programming constructs such as conditionals, loops, and functions which correspond to other popular
programming languages such as C++ or Python. MATLAB is a commercial product, which can be prohibitively
expensive for students to acquire, however the College of Engineering at Ohio State provides all engineering students
access to MATLAB under a site license.

2. Course Overview

The course is designed as a 10-week class (1 quarter) which meets four days a week: three lecture classes and
one lab meeting. The textbook covers both introductory programming skills and techniques specific to MATLAB for
vector and matrix manipulation [3].

2.1. Topics
Students are first given an informal treatment of algorithms and how to view problem solving as a set of steps.
Students are then exposed to the following concepts in order:
Expressions: Basic arithmetic, logical, and boolean expressions. Constructing complex expressions from smaller,
simpler ones.
Variables: Using variables in expressions and changing their values with assignment.
Vectors and Matrices: Vectors as a first-class datatype, vector creation, vector operations (e.g. sum(), max()
etc.), and expressions with vectors (e.g. vector-vector operations, vector-scalar operations). Element-wise operations:
updating an element, slicing, using simple and complex expressions to index into vectors. Extend same concepts to
matrices.
Visualization: Displaying data, input/output, 2D and 3D plotting.
 D.B. Larkins, W. Harvey / Procedia Computer Science 1 (2012) 913–919
D. B. Larkins et al. / Procedia Computer Science 00 (2010) 1–7
915
3

Functions: Functional abstraction, input parameters, return values, returning multiple values.
Conditionals: Introduce if statements, work with control-flow diagrams to differentiate between if, if...else,
and if...elseif conditional blocks.
Iteration: Bounded definite loops using for, work examples demonstrating iteration space, variable loop bounds,
etc.
Composition: Whole program construction, composition of various program elements (conditionals, loops, functions,
etc.)

2.2. Course Structure

Students are assigned homework every week to assess their level of understanding the lecture material. The
dedicated lab session is used as a closed laboratory at the beginning of the course, with assignments due at the end of
class. The closed lab sessions focus on giving students hands-on experience with the lecture topics described above.
The closed labs are designed to develop basic programming skills and the direct application of MATLAB-specific
syntax and constructs.
As students gain more sophisticated programming knowledge and are able to compose rudimentary programs,
the lab session is used to work on four in-depth laboratory assignments. These labs are more focused on problem
definition and specification, leaving students to develop their own algorithms and compose programming elements to
generate the correct output. The first two laboratory assignments introduce students to the problem solving process by
asking them to develop solutions to conceptually straightforward problems. The first focuses on manually computing
basic statistical parameters of a predefined data set and the second requires the use of conditionals to implement a
decision tree.
The last two laboratories introduce students to basic image processing theory and assign students the task of using
this knowledge to detect the edges in an image. These assignments are discussed in detail below.

3. Understanding Computational Science through Image Processing

3.1. Why Image Processing?

Since this course is an introductory course, students are assumed to understand basic calculus concepts, but may
not have had linear algebra, signal processing, partial differential equations, etc. Any assignments designed to apply
programming knowledge and experience to a computational problem must either be readily understood with only
a rudimentary calculus background, or the underlying theory must be introduced in class. Furthermore, given that
student demographics range over different science and engineering disciplines, no one application will be relevant for
all students.
Image processing is a good choice for understanding computational science for several reasons. First, it is a
challenging subject that is itself the subject of active research and based on mathematical and statistics theory common
to many fields [4]. Edge detection, in particular, is a useful pedagogical tool because although the underlying theory
can be complex, it can be understood intuitively with an understanding of basic calculus. This intuitive approach to
edge detection is bolstered by the rich support for image processing provided by MATLAB, allowing a great deal
of flexibility in deciding which aspects of the problem are exposed to students and which can be handled by the
computational environment. Lastly, image processing is inherently visual. Students are working with images which
can be viewed at every step of the program, strengthening the intuitive sense of process and algorithmic thinking.

3.2. Overview of Edge Detection

Given a digital representation of a grayscale image, the objective is to identify pixels that correspond to edges in
the image. For example, Figure 1 illustrates a sample input image and detected edge pixels.
Mathematically, a grayscale image can be represented via the map

I : R2 → R (1)
This representation provides an intuitive interpretation of an image as a landscape where the intensity at a point
is equivalent to its height. Thus, edge-like structures in the image become barriers and cliffs where the height of the
916 D.B. D.
Larkins, W. Harvey / Procedia Computer Science 1 (2012) 913–919
B. Larkins et al. / Procedia Computer Science 00 (2010) 1–7 4

(a) Original image (b) Gradient in x direction (c) Gradient in y direction (d) Gradient magnitude

Figure 1: Edge detection framework in which the gradient magnitude is used to identify pixels that correspond to edge features in the input image.

landscape changes abruptly. These features can be characterized as collections of points where (locally) the change in
height in one direction is large, yet the change in height in an orthogonal direction is small.
For every point x ∈ R2 , the rate of change of I is determined by the gradient, defined as follows:

(∇I) : R2 → R2
(2)
x → (I x (x), Iy (x))

∇I(x) captures the direction in which the rate of increase of I at x is maximized, with ∇I(x) representing the mag-
nitude of change. Thus, applying a threshold to ∇I(x) provides a simple classifier that identifies regions belonging
to the edge features in the image.
While this mathematical framework is sound, its implementation requires adaptation to work reliably with digital
grayscale images. An 8-bit grayscale digital image of dimensions (w × h) can be represented using the map

L : [1, w] × [1, h] → [0, 255] (3)

The partial derivatives of L along the x and y directions can be computed as follows:

L x (x, y) = L(x + 1, y) − L(x, y) (4)

Ly (x, y) = L(x, y + 1) − L(x, y) (5)

However, due to discretization artifacts and presence of noise, it is beneficial to locally smooth L and to use a
small neighborhood around each point to estimate the partial derivatives. This leads to the Sobel operator formulation
for estimating the partial derivates:
⎡ ⎤ ⎡ ⎤
⎢⎢⎢−1 0 +1⎥⎥⎥ ⎢⎢⎢+1 +2 +1⎥⎥⎥
⎢⎢⎢ ⎥⎥⎥ ⎢⎢⎢ ⎥
L x = ⎢⎢−2 0 +2⎥⎥ ∗ L Ly = ⎢⎢ 0 0 0 ⎥⎥⎥⎥ ∗ L (6)
⎣ ⎦ ⎣ ⎦
−1 0 +1 −1 −2 −1
where ∗ denotes discrete convolution. The result of using the Sobel operator [4] to compute the partial derivatives of
an input image is shown in Figures 1(b) and 1(c). When computing the partial derivatives along the boundaries of L,
the image intensities beyond the borders are considered to be zero, although other heuristics can be used. It is possible
that the images L→ →
x and Ly of L x and Ly contain values beyond the range of valid 8-bit grayscale quantities. In this
case, they can be corrected using the following affine transformation:

255(L x − min L→
x )
255(Ly − min Ly→ )
Lˆx = Lˆy = (7)
max L→x − min L →
x max Ly→ − min Ly→

With Lˆx and Lˆy in hand, the gradient magnitude G(x) is computed using the L2 -norm, and edge points are detected
by applying a threshold τ ∈ [0, 255] to G.
 D.B. Larkins, W. Harvey / Procedia Computer Science 1 (2012) 913–919
D. B. Larkins et al. / Procedia Computer Science 00 (2010) 1–7
917
5

3.3. Exercises
Students work with the edge detection problem in two lab assignments over three weeks. The first lab assignment
is primarily focused on introducing students to working with image data, creating functions, and using basic loop
constructs. The second lab focuses more on the implementation of basic image processing algorithms.
First Lab: This project consists of two functions and a main program which perform edge detection on an image.
1. Create a function to display a matrix as an image.
This function will be used to display progress throughout the lab, and also can aid in debugging. Functions
from the MATLAB Image Processing Toolbox are called to do the actual display work, having students focus
on function creation and parameter passing.
2. Create a function to convert a grayscale image to a monochrome image at some threshold.
This function takes a grayscale image matrix and a threshold value, t, and produces a new monochrome image
matrix. Students implement a function which iterates over each point, p, in the grayscale image and sets the
corresponding point, p , in the monochrome image to either black or white depending on whether or not the
intensity value of p is greater than the threshold, t. Students are required to implement this with an explicit loop
structure which requires the composition of loop and conditional constructs.
The main program uses these two functions to implement the edge detection process:
1. Read an image file into a matrix variable.
This functionality is provided by the Image Processing Toolbox.
2. Create gradient filter masks in the x and y directions.
This step creates the two 3 × 3 Sobel operator filters which are used to detect edges in each of the x and
y directions. These filters are created from Image Processing Toolbox functions but could be defined in the
exercise or derived from theory. There are a number of filters and techniques used to detect edges in an image
which could be used, such as Sobel, Canny, Laplacian, or Prewitt filters[5, 6, 4].
3. Apply the gradient filters to the image.
In this exercise, students call the MATLAB function, imfilter(), which performs the discrete convolution
of the filter and the image matrix. The resulting matrix consists of intensity values which represent the rate
of change of the intensity values in the source image. The second edge detection laboratory has the students
implement the filtering convolution, without relying on the imfilter() function.
4. Use MATLAB operations to compute gradient magnitude image.
The magnitude is computed by taking the x and y filtered gradient images and computing a new magnitude
image, which is defined by: m = x2 + y2 .
5. Call threshold function to convert grayscale magnitude image to a monochrome image.
The grayscale magnitude image is converted to a monochrome image by the student-authored threshold func-
tion. Students iterate over a range of threshold values and determine which ones eliminate artifacts from noise
without sacrificing too much edge detail.
Second Lab: For the second image processing assignment, students implement the discrete convolution operator
which performs the filtering operation. Students can assume that the filter mask is of a fixed size, but extra credit was
offered for implementations which operate on arbitrarily-sized masks.
The basic structure of this exercise is similar to the first, with the primary difference being the student-developed
filter function. Students also use this filtering operation for an explicit smoothing step using a Gaussian smoothing
mask. The filtering operation can be thought of as a stencil operation on each element in the input image.
The image filtering function performs three key tasks:
1. Output image padding: Since the filtering operation is similar to a stencil computation, the original image
must be zero-padded in order to ensure well-defined output in the filtered image. Students must determine the
appropriate amount of padding based on the filter size and add a corresponding border of zero-valued elements.
2. Discrete convolution operation: The convolution operation involves iterating over each element in the source
image and taking the sum of the products of neighboring elements with the filter values. This requires students
to write a nested loop statement with the computation using fixed offset indices of the loop index variables.
918 D.B. D.
Larkins, W. Harvey / Procedia Computer Science 1 (2012) 913–919
B. Larkins et al. / Procedia Computer Science 00 (2010) 1–7 6

3. Normalization: The convolved image matrix may contain values which are outside the 8-bit intensity value
range from 0 to 255, including negative values. Students must scale the values in the filtered image to the legal
range, again giving them experience with problem-motivated matrix transformations.

4. Experience and Observations

Reviewing the lab submissions for both image processing exercises illuminated some key insights as to how
students were able to integrate the different program elements to construct a complete solution. For example, the
minimal solution to implement the discrete convolution operator used in image filtering requires two nested loops, to
iterate over each point in the source image. The filtering operation was most commonly implemented by explicitly
summing the products of a point in the filter matrix with a point in the image matrix. However, several students
implemented the filtering operation with the following form:
For a 3 × 3 filter, f and a source image, im:

for i = 1,rows
for j = 1,cols
out(i,j) = sum(f(1,1:3).*im(i-1,j-1:j+1)) + ...
end
end

This approach blends both the benefits of using a high-level vector operations such as sum() with traditional
computational structures like loops. Indeed, these skills are exactly what we want students to take away from the
course: If high-level functions or language level support can solve the problem at hand they are preferred, otherwise,
use lower-level techniques to implement the algorithm needed.

4.1. Student Survey Results

At the end of the course, students were surveyed and reported on their experience with different aspects of the
course and on the use of the edge detection exercises in particular. The questions about the image processing compo-
nents were whether they felt like edge detection was understandable given their math background, they felt like they
learned from the experience, if it changed their opinion of computing and using MATLAB, and if so whether it was a
positive or negative change. These results are summarized in Table 1.
Overall, over 81% of the students felt that the image processing concepts related to edge detection were under-
standable given their math and science background. This fits with our goal to pick a problem which is approachable
to students with a basic calculus background. This also indicates that some students need some more context prior
to undertaking the assignment. Future classes will spend more time on the math concepts involved and how they are
related to the edge detection problem.
Over 74% of students felt like they learned from the edge detection laboratories. In the department course feedback
forms, some students stated that they didn’t feel that the edge detection labs were relevant to their chosen discipline.
While this is certainly the case for many students, it will always be difficult to pick a problem of relevance for all
students given their diverse intended fields. Just under 89% of students changed their opinion of computing and the
use of MATLAB as a computational tool. Of these students, 83% came to view the practice of programming and
using MATLAB to solve problems in a more favorable light.

5. Conclusion

Introducing engineering students to computational science is an important task early in the curriculum. Early ex-
posure to computational tools allows students to use them throughout their education and into their vocation. Exposing
students to computational science early, however, implies minimal math and science prerequisites which impacts the
topics and applications covered in the course.
We have developed coursework and lab exercises which use image processing as a model application domain to
give students experience solving scientific problems with modern high-level computational tools. This course has
 D.B. Larkins, W. Harvey / Procedia Computer Science 1 (2012) 913–919
D. B. Larkins et al. / Procedia Computer Science 00 (2010) 1–7
919
7

Question Response Count

not at all 2
1. How understandable is the concept of edge detection in an image difficult to understand 3
given your math/science background? understandable 17
straightforward 5
not at all 1
2. How much do you feel like you learned through the experience
not very much 6
of solving the edge detection problem as a real-world application of
somewhat 11
MATLAB?
a lot 9
3. Did your experience with the edge detection labs change your yes, a lot 8
opinion of computing and the use of MATLAB to solve real-world yes, a little 16
problems? no, not at all 3
4. If you answered yes, how did your opinion change? Do you now more favorably 20
view programming and MATLAB more or less favorably? less favorably 4

Table 1: Summary of Student Survey Data. (n = 27)

been successfully piloted at Ohio State and has been found to be a useful tool in engaging students in computational
science and helping them learn how to solve scientific problems.
Much of the course materials (slides, laboratory assignments, homework) are available at http://www.cse.
ohio-state.edu/~larkins/cse294p.

References
[1] G. J. Borse, Numerical Methods with MATLAB: A Resource for Scientists and Engineers, International Thomson Publishing, 1996.
[2] J. Kepner, Parallel MATLAB for Multicore and Multinode Computers, Society for Industrial and Applied Mathematics (SIAM), 2009.
[3] D. T. Kaplan, Introduction to Scientific Computation and Programming, Brooks/Cole, 2004.
[4] D. A. Forsyth, J. Ponce, Computer Vision: A Modern Approach, Prentice Hall Professional Technical Reference, 2002.
[5] J. Canny, A computational approach to edge detection, IEEE Trans. Pattern Anal. Mach. Intell. 8 (6) (1986) 679–698.
[6] J. Canny, Finding edges and lines in images, in: MIT AI-TR, 1983.