1. Image segmentation and detection. 2. Feature calculation. 3. Object tracking.
Budget: £70 – £100 GBP
Task 1: Image Segmentation and Detection
Download and unzip the file ‘skin lesion dataset.zip’ from Blackboard. You should obtain a set of
120 images. Among those images, there are 60 skin lesion colour images and 60 corresponding
binary masks (ground-truth segmentation).
Please use image processing techniques to implement the following tasks. Please note that you
are encouraged to develop one model with same parameter settings for all the images.
Task 1: Object segmentation. For each skin lesion image, please use image processing
techniques to automatically segment lesion object. Examples of the lesion image (Fig.1(a) and
the segmented lesion
Task 2: Segmentation evaluation. For each skin lesion image, calculate the Dice Similarity Score
(DS) which is defined in Equation 1; where M is the segmented lesion mask obtained from Task
1, and S is the corresponding ground-truth binary mask.
DS =
2|?⋂?|
|?|+|?|
(1)
The calculated DS shall be between 0 and 1. For example, DS is 1 if your segmentation matches
perfectly with the ground-truth mask, whist DS is 0 if there is no overlap between your
segmentation and ground-truth mask.
Your report should include: 1) For three skin images (ISIC_0000019, ISIC_0000095 and
ISIC_0000214), you are required to put the original images, final segmented lesion binary images,
the calculated DS value for each of the three images; 2) for all the 60 skin images, please provide
a bar graph with x-axis representing the number of the image, and y-axis representing the
corresponding DS. 3) Calculate the mean and standard deviation of the DS for all the 60 images.
4) briefly describe and justify the implementation steps.
Task 2: Feature Calculation
Download the Image (‘ImgPIA.jpeg’) from Blackboard. This part of the assignment will deal with
the area of Feature Extraction, in both the Frequency and Spatial domains.
Task 1: Read the image (‘ImgPIA.jpeg’), and select the features for both radius and direction as
described in the Spectral Approach session of the Feature Extraction lecture. For additional
marks you can change the values of radius and angle, and present those values in a plot or table.
Task 2: Read the image (ImgPIA.jpeg), and select features from the image histogram (i.e. 1st
order), at least six (6) features from the co-occurance matrix (the original paper by Haralick has
also made available to you), and at least five (5) features from the Gray Level Run Length (GLRL)
matrix. Please note that both the co-occurance and GLRL based features can be directional and
as a function of distance between pixel co-ordinates. For additional marks you can change the
bit-depth of the image (i.e. 8, 6, 4 bit), and recalculate the features presenting them as a plot or
table.
For both tasks analysis and discussion of your findings is expected.
Task 3: Object Tracking
Download from Blackboard the data files 'x.csv' and 'y.csv', which contain the real coordinates
[x,y] of a moving target, and the files 'a.csv' and 'b.csv', which contain their noisy version [a,b]
provided by a generic video detector (e.g. frame-to-frame image segmentation of the target).
Implement a Kalman filter with a software application that accepts as input the noisy
coordinates [a,b] and produces as output the estimated coordinates [x*,y*]. For this, you should
use a Constant Velocity motion model F with constant time intervals Δt = 0.1 and a Cartesian
observation model H. The covariance matrices Q and R of the respective noises are the
following:
? = [
0.16 0 0 0
0 0.36 0 0
0 0 0.16 0
0 0 0 0.36] ? = [
0.25 0
0 0.25]
1) You should plot the estimated trajectory of coordinates [x*,y*], together with the real [x,y]
and the noisy ones [a,b] for comparison.
2) You should also assess the quality of the tracking by calculating the mean and standard
deviation of the absolute error and the Root Mean Squared error (i.e. compare both noisy and
estimated coordinates to the ground truth).
Download and unzip the file ‘skin lesion dataset.zip’ from Blackboard. You should obtain a set of
120 images. Among those images, there are 60 skin lesion colour images and 60 corresponding
binary masks (ground-truth segmentation).
Please use image processing techniques to implement the following tasks. Please note that you
are encouraged to develop one model with same parameter settings for all the images.
Task 1: Object segmentation. For each skin lesion image, please use image processing
techniques to automatically segment lesion object. Examples of the lesion image (Fig.1(a) and
the segmented lesion
Task 2: Segmentation evaluation. For each skin lesion image, calculate the Dice Similarity Score
(DS) which is defined in Equation 1; where M is the segmented lesion mask obtained from Task
1, and S is the corresponding ground-truth binary mask.
DS =
2|?⋂?|
|?|+|?|
(1)
The calculated DS shall be between 0 and 1. For example, DS is 1 if your segmentation matches
perfectly with the ground-truth mask, whist DS is 0 if there is no overlap between your
segmentation and ground-truth mask.
Your report should include: 1) For three skin images (ISIC_0000019, ISIC_0000095 and
ISIC_0000214), you are required to put the original images, final segmented lesion binary images,
the calculated DS value for each of the three images; 2) for all the 60 skin images, please provide
a bar graph with x-axis representing the number of the image, and y-axis representing the
corresponding DS. 3) Calculate the mean and standard deviation of the DS for all the 60 images.
4) briefly describe and justify the implementation steps.
Task 2: Feature Calculation
Download the Image (‘ImgPIA.jpeg’) from Blackboard. This part of the assignment will deal with
the area of Feature Extraction, in both the Frequency and Spatial domains.
Task 1: Read the image (‘ImgPIA.jpeg’), and select the features for both radius and direction as
described in the Spectral Approach session of the Feature Extraction lecture. For additional
marks you can change the values of radius and angle, and present those values in a plot or table.
Task 2: Read the image (ImgPIA.jpeg), and select features from the image histogram (i.e. 1st
order), at least six (6) features from the co-occurance matrix (the original paper by Haralick has
also made available to you), and at least five (5) features from the Gray Level Run Length (GLRL)
matrix. Please note that both the co-occurance and GLRL based features can be directional and
as a function of distance between pixel co-ordinates. For additional marks you can change the
bit-depth of the image (i.e. 8, 6, 4 bit), and recalculate the features presenting them as a plot or
table.
For both tasks analysis and discussion of your findings is expected.
Task 3: Object Tracking
Download from Blackboard the data files 'x.csv' and 'y.csv', which contain the real coordinates
[x,y] of a moving target, and the files 'a.csv' and 'b.csv', which contain their noisy version [a,b]
provided by a generic video detector (e.g. frame-to-frame image segmentation of the target).
Implement a Kalman filter with a software application that accepts as input the noisy
coordinates [a,b] and produces as output the estimated coordinates [x*,y*]. For this, you should
use a Constant Velocity motion model F with constant time intervals Δt = 0.1 and a Cartesian
observation model H. The covariance matrices Q and R of the respective noises are the
following:
? = [
0.16 0 0 0
0 0.36 0 0
0 0 0.16 0
0 0 0 0.36] ? = [
0.25 0
0 0.25]
1) You should plot the estimated trajectory of coordinates [x*,y*], together with the real [x,y]
and the noisy ones [a,b] for comparison.
2) You should also assess the quality of the tracking by calculating the mean and standard
deviation of the absolute error and the Root Mean Squared error (i.e. compare both noisy and
estimated coordinates to the ground truth).
Related categories:
Computer Vision