辅导algorithm留学生、Java编程设计讲解、辅导Python/C++、讲解retina image
- 首页 >> 其他 Coursework 2018-2019
Task Description and Guidelines
Segment the retina blood vessel. (16 marks)
Retina blood vessel segmentation is challenging. Given the image on the left, it is very difficult
to accurately segment the blood vessel as shown on the right, especially the thin ones. You are
free to design your own algorithm to segment the blood vessel, down to per pixel level. NOTE:
You are not required to achieve the state-of-the-art results, but you should achieve
satisfactory results.
Requirements & Guidelines
40 retina images are provided, together with the label images indicating where the blood
vessel is, and the mask image masking out the background of retina image.
You should test your algorithm on retina_images/1.tif only, and show the segmentation
results for this image. You should never use label_images/1.tif when you design your
algorithm. All other images being provided can be used as training images when you design
your algorithm. When test your algorithm, you should compare the segmentation results of your algorithm
with the ground-truth image (label_images/1.tif). You should report the following:
The percentage of blood vessel pixels that is being correctly classified as blood vessel.
Denoted as P.
The percentage of background pixels (only consider the region in the mask.) that is being
correctly classified as background. Denoted as N.
The percentages of pixels are being correctly classified. (only consider the region in the
mask). Denoted as T.
For example, assume in the mask we have 100 pixels, 10 pixels are retina and 90 pixels
are background. The segmentation result is that 12 pixels are retina (only 8 are true retina
and 4 are false retina) and 88 pixels are background. (only 86 are true background and 2
are false background.) Then P = 8/10 = 80%; N = 86/90 = 95.56%; T = (8+86)/100=94%.
You may consider the following filters: (It is up to you to choose which one to use and how
many you may use. You may use other filters if appropriate. You may use FILTERS from
internet resources, with proper acknowledgement.
Laplacian of Gaussian filter.
Difference of Gaussian filter.
Canny filter.
Match filter.
Gabor filters.
You may directly process the filtered images to obtain the per-pixel classification results of
the retina image. Or you may try to combine the results from several filters, e.g. to treat the
filtered output of other images as the training samples, and train a per-pixel classifier to
classify the test image retina_images/1.tif.
You should implement one main function that calls other functions to generate all the results,
e.g. task3main().Marking Criteria
You should include the explanation of your algorithm, the intermediate results and the final
results in your report. Marks will be given by considering:
The quality of your per-pixel level segmentation results for the first image. You should try to
maximize both P and N. (6 marks)
The explanation and justification of your algorithm. (6 marks)
Discuss advantages and disadvantages of your algorithm. (2 marks)
Whether your codes are well documented, including comments, and explanation of variables
at the beginning of each function. (2 marks)
Report
All images on the report should be large enough for visual inspection of the image quality.
How to submit
You should zip all the matlab files, the result images (only upload the segmentation result of
the test image) and the PDF file for the report in ONE zipped file.
Task Description and Guidelines
Segment the retina blood vessel. (16 marks)
Retina blood vessel segmentation is challenging. Given the image on the left, it is very difficult
to accurately segment the blood vessel as shown on the right, especially the thin ones. You are
free to design your own algorithm to segment the blood vessel, down to per pixel level. NOTE:
You are not required to achieve the state-of-the-art results, but you should achieve
satisfactory results.
Requirements & Guidelines
40 retina images are provided, together with the label images indicating where the blood
vessel is, and the mask image masking out the background of retina image.
You should test your algorithm on retina_images/1.tif only, and show the segmentation
results for this image. You should never use label_images/1.tif when you design your
algorithm. All other images being provided can be used as training images when you design
your algorithm. When test your algorithm, you should compare the segmentation results of your algorithm
with the ground-truth image (label_images/1.tif). You should report the following:
The percentage of blood vessel pixels that is being correctly classified as blood vessel.
Denoted as P.
The percentage of background pixels (only consider the region in the mask.) that is being
correctly classified as background. Denoted as N.
The percentages of pixels are being correctly classified. (only consider the region in the
mask). Denoted as T.
For example, assume in the mask we have 100 pixels, 10 pixels are retina and 90 pixels
are background. The segmentation result is that 12 pixels are retina (only 8 are true retina
and 4 are false retina) and 88 pixels are background. (only 86 are true background and 2
are false background.) Then P = 8/10 = 80%; N = 86/90 = 95.56%; T = (8+86)/100=94%.
You may consider the following filters: (It is up to you to choose which one to use and how
many you may use. You may use other filters if appropriate. You may use FILTERS from
internet resources, with proper acknowledgement.
Laplacian of Gaussian filter.
Difference of Gaussian filter.
Canny filter.
Match filter.
Gabor filters.
You may directly process the filtered images to obtain the per-pixel classification results of
the retina image. Or you may try to combine the results from several filters, e.g. to treat the
filtered output of other images as the training samples, and train a per-pixel classifier to
classify the test image retina_images/1.tif.
You should implement one main function that calls other functions to generate all the results,
e.g. task3main().Marking Criteria
You should include the explanation of your algorithm, the intermediate results and the final
results in your report. Marks will be given by considering:
The quality of your per-pixel level segmentation results for the first image. You should try to
maximize both P and N. (6 marks)
The explanation and justification of your algorithm. (6 marks)
Discuss advantages and disadvantages of your algorithm. (2 marks)
Whether your codes are well documented, including comments, and explanation of variables
at the beginning of each function. (2 marks)
Report
All images on the report should be large enough for visual inspection of the image quality.
How to submit
You should zip all the matlab files, the result images (only upload the segmentation result of
the test image) and the PDF file for the report in ONE zipped file.