联系我们： 手动添加方式: 微信>添加朋友>企业微信联系人>13262280223 或者 QQ: 1483266981

EEEM063/LSA/18/19/pgs6/Req: None
UNIVERSITY OF SURREYc
Faculty of Engineering and Physical Sciences
Department of Electronic Engineering
Undergraduate Programmes in Electronic Engineering
Module EEEM063; 15 Credits
Computer Vision and Pattern Recognition
FHEQ LEVEL 6 Examination
Time allowed: Two hours Semester 2 2018/19
READ THESE INSTRUCTIONS
Answer only THREE questions
If you answer more than THREE questions, the highest scoring THREE questions will
count towards this assessment.
On the front sheet of each book, complete a list of questions attempted in that book in the
order they appear. Then draw a line below this list and add the question numbers you have
attempted in the other books.
Where appropriate the mark carried by an individual part
of a question is indicated in square brackets [ ].
Additional materials
Candidates may use only calculators which are non-programmable and with no
alphanumeric memory.
c
Please note that this exam paper is copyright of the University of Surrey and may not be
reproduced, republished or redistributed without written permission.
EEEM063/LSA/18/19/pgs6/Req: None
1.
(a) Explain what is meant by a gradient in the context of the backpropagation algorithm
used to train a neural network.
[10%]
(b) Consider the simple multilayer perceptron (MLP) neural network below. Only the
final two layers of the network are shown. Assume the weights {w1, w2, w3, w4}
have values {0.4, 0.5, 0.6, 0.7} respectively. There are no biases.
(i) An input is presented to the MLP resulting in hidden layers h1 and h2 firing with
activations outh1 = 0.2 and outh2 = 0.3 respectively. Complete the remainder of the
feed-forward pass of the network to calculate the outputs outo1 and outo2. Recall
that the equation for the Sigmoid (logit) function is
σ(x) = 1
1 + e
x
(1)
[10%]
(ii) Use your solution to part (b,i) to calculate the mean squared error (MSE) total loss
of this network, given a one-hot target for o1.
[10%]
(iii) Use your solution to part (b,ii) to calculate the update on weight w1 due to the
backpropagation algorithm.
[20%]
(c) What problems are typically encountered with gradients when training very deep
neural networks.
Page 2 SEE NEXT PAGE
EEEM063/LSA/18/19/pgs6/Req: None
[10%]
(d) Explain how the following deep convolution neural networks (CNNs) mitigate
problems with gradients during training.
(i) VGG-19 network
[10%]
(ii) GoogLeNet Inception network
[10%]
(iii) CNNs typically require a large amount of training data in order to learn effective
models. Briefly outline TWO strategies for working around this issue when training
a CNN with limited data.
[20%]
Page 3 SEE NEXT PAGE
EEEM063/LSA/18/19/pgs6/Req: None
2.
(a) Digital image warping can be performed using a forward mapping or backward
mapping method. With the aid of diagrams compare and contrast the operation of
these two methods, and state one advantage and disadvantage of each.
[20%]
(b) Consider the following RGB colour image
(i) Perform bi-linear interpolation to compute the RGB colour at (6.2,2.7)
[20%]
(ii) Perform Gaussian interpolation to compute the RGB colour at (6.4,2.6). Use the
equation for a Gaussian with standard deviation σ = 1:
G(x) = 1
√
2π
e

2
x
2
(2)
[20%]
(c) Convert the image shown in part (b) into the HSV colour space.
[20%]
(d) Can the RGB colour space represent all possible colours Briefly outline the design
of an experiment that could prove your answer.
[20%]
Page 4 SEE NEXT PAGE
EEEM063/LSA/18/19/pgs6/Req: None
3.
(a) Explain what is meant by an encoder-decoder convolutional neural network (CNN)
and state the type and function of layers typically used within such networks.
[10%]
(b) Sketch the architecture of an CNN auto-encoder that could be trained to de-noise
an image. Describe the training methodology for such a network including any prior
that could be used to regularize its output.
[20%]
(c) Explain with the aid of a diagram what is meant by a Deep Image Prior (DIP). How
can the DIP constrain image reconstruction
[20%]
(d)
(i) Explain the aperture problem within the context of optical flow (motion) estimation.
[10%]
(ii) Write down the brightness constancy equation for optical flow and develop the mathematics to show how the aperture problem is manifested within it.
[15%]
(iii) Sketch the architecture of a CNN encoder-decoder for learning optical flow and
briefly explain how it could be trained.
[15%]
(iv) How could appearance and motion information be combined within a single CNN
for the purpose of activity recognition
[10%]
Page 5 SEE NEXT PAGE
EEEM063/LSA/18/19/pgs6/Req: None
4.
(a) Describe the key stages of the AlexNet convolutional neural network (CNN) architecture for object recognition, as defined in [Krizhevsky et al., NIPS 2012]. You
do not need to provide numerical parameters for each layer e.g. window sizes, but
should state the name, order and function of the CNN layers in the network.
[20%]
(b) Describe process of Activation Maximisation (AM) and how it may be used to
uncover bias in data used to train an object recognition CNN such as Alexnet.
[20%]
(c) How could AM be adapted to create an adversarial image capable of causing a CNN
to misclassify an image.
[20%]
(d) Describe a simple method of disentangling the content and the style of an image
using a CNN.
[20%]
(e) Using your answer to part (d) or otherwise, explain how could AM be adapted to
transfer the style of one image onto the content of another.
[20%]
Internal Examiner: Dr J. Collomosse
External Examiner: Prof. N. Canagarajah
Page 6 FINAL PAGE