Active-contour-fitting

Task: Active Contour Fitting

In addition to implementing active contours, the following image manipulation tasks are done using OpenCV and Python:

Allow the user to specify the initial position of the input curve though mouse interaction
Implement the greedy algorithm for minimizing the energy functional
Show the evolution of active contour until convergence
Interactively controlled parameters should include: time delay between iterations, continuity/curvature/external energy coefficients, and the fraction of points that has to reach a local minimum to determine convergence
Display the results overlaid in color over the original grayscale image

Proposed Solution:

E(f(s)) = sum(alpha*continuity_energy + beta*curvature_energy + gamma*external_energy)

Start with a parametric curve, and begin shrinking it as we greedily minimize the total energy of the image
Gaussian filter(sigma=3) has been used to perform smoothing

Implementation Details

The initial parametric curve chosen is a circle
The image window displayed has trackbars for setting the values of: centre of the circle, radius of the circle, alpha, beta, gamma, fraction of points to reach local minima before convergence, time delay (in ms) between iterations and Ready.
Adjust these values as required and press any key (except ‘q’) to begin fitting the contour.
In order to begin fitting the contour, make sure that the ‘Ready’ trackbar is switched to 1. Contour fitting will NOT begin otherwise.
Evolution of contour can be seen 10 iterations at a time, once the contour fitting begins. The current iteration can be seen at the top
At any point during the program, pressing the key ‘q’ will terminate the currently running operation. For example, if ‘q’ is pressed while the iterations are being performed, the iterations stop and the most recently detected contour is displayed.
In the final result displayed, the dashed-circle (in red) is the initial circle specified by the user. The contour is displayed by a blue-edged polygon.
2 images have been used in this report to showcase the fitting process. Starting with a simpler image (of a cup), it is easier to visualize the process of fitting the contour. The second image is relatively complex (Lenna), to show the robust working of the algorithm.

Results:

Image 1: Cup

Parameters: alpha=0.05, beta=0.05, gamma=0.01, no. of iterations=1000

Original image, initially selected region for contour fitting

Evolution of contour every 100 iterations

Final result after 1000 iterations

Image 2: Lenna

Parameters: alpha=0.05, beta=0.05, gamma=0.01, no. of iterations=1000

Original image, initially selected region for contour fitting

Evolution of contour every 100 iterations

Final Result after 1000 iterations

Inferences:

As it can be seen from the two examples of contour fitting shown above, the contours start to fit at the expected points in the selected region.
It is more evident in the example of the cup, how the contours evolve to completely border the cup.
In the Lenna example, it can be seen that the contour has started to take the shape of the hat and the edges in the background.
Subsequent iterations show that we can completely fit the contour as per the selected region in the image
Finally, the user can experiment and see the effects of the different energy coefficients in the contour obtained as a result.

References:

Perform Fitting using Active Contours to detect edges of an object in an image.