douglasteeple / ae0732

Aeolus Project

AE0732 Programming Excercise

This excercise implements both a class with the methods given in the AE0732 project definition and a command line driver to excercise the methods with test images.

Class ImageAnalysisService

1. FIND_REGION accepts an RGB image, an (X, Y) pixel location, and a distance parameter to control a greedy style algorithm that "flood fills" the nearby pixels in order to find a contiguous patch of pixels that are similar in color. The return value is a vector of cv::Points. The RGB image is converted to HSV to easily select by color (H). The maximum distance from the color at point x,y is given as a parameter. The hue at the x,y point is optionally returned in the "*hue" argument.

2. FIND_PERIMETER accepts the output of FIND_REGION and outputs contiguous pixels that border the region. These pixels are a subset of pixels returned by FIND_REGION.

3. FIND_SMOOTH_PERIMETER accepts the output of FIND_REGION and outputs contiguous pixels that border the region. These pixels are a subset of pixels returned by FIND_REGION. The perimeter values are smoothed to remove jagged contours, small artifacts.

   Smoothing:

   • removes jagged contours, small artifacts, etc.

   • remains generally close to the non-smoothed original boundary

   • potentially expressed via a lower degree parametric representation of the smoothed perimeter, such as splines or bezier curves.

   A number of smoothing options are available, as defined in the smoothing_type enum as: "Blur, MedianBlur, GuassianBlur, BilateralFilter".

4. DISPLAY_IMAGE provides a method for displaying loaded RGB images.

5. DISPLAY_PIXELS displays an image representation of the results from the above FIND_* operators in a region point vector.

6. SAVE_PIXELS accepts the region output saves as an image to a file.

7. SAVE_IMAGE saves and image to a named file.

Test Driver

The test driver in main.cpp accepts as arguments, the image file path, and output file path, x,y coordinates in the image from which the initial hue value is taken and optionally an angular distance (arbitrary values from 0 to 179) in hue from which the point value will be considered to be in the region.

Build

cd to directory

cmake .

make

Run Example

./ImageAnalysisService AeolusTestImages/test1.png 10 20

Smoothing:

./ImageAnalysisService AeolusTestImages/test3.png 100 100

To run all test cases:

./run.sh

Sample Results

The image test4.jpg gave best results with MedianBlur and a kernel size of 31.

./ImageAnalysisService AeolusTestImages/test4.jpg 2200 600

Original Image	Region Image
Perimeter	Smoothed Perimeter

The image test3.png gave best results with MedianBlur and a kernel size of 5.

../ImageAnalysisService AeolusTestImages/test3.png 100 100

Original Image	Region Image
Perimeter	Smoothed Perimeter

DISCUSSION

I opted to simply offer a number of smoothing options to the user. The best results I found came from one of the simplest: median smoothing. It removes the spot noise, but has the unfortunate side effect of rounding corners of the perimeter. It does a better job on more complex images such as test4 when the kernel size is large (31), and better on 2 or 3 color images such as test3 with a smaller kernel size of 5. In both cases Median Smoothing gave the best results.

Smoothing noise from the original image (not just the point vector) could improve the results.

The main driver could be improved with more command line user options. I kept it simple - producing output from each step.

TODO

• better perimeter smoothing support