IMAGE_ASS3

                                  Hacettepe University
                  Computer Engineering - Artificial Intelligence Engineering


                     AIN 432 – Fundamentals of Image Processing Lab.
                                   Assignment 3 Report

               Image Segmentation using RGB and SuperPixel Future with KMeans

Drive Link: https://drive.google.com/drive/folders/19JcIOSNU3NeVUF4ZTwtXjkWr2Th6RoLQ?usp=sharing

Metehan Sarikaya 21993049

30.12.2023
Problem: We will segment images based on 5 different features which are; RGB color at pixel-level, RGB color and location feature at pixel-level, mean RGB feature at superpixel level, RGB color histogram at superpixel level, mean Gabor response at superpixel level using KMeans Clustering algorithm and do experiments with k and segments value. We will do our experiment on 5 different image and examine the results.

Methodology: We will use K-means clustering algorithm for image segmentation by using pixel level and superpixel representation of an input image. In Pixel-Level Feature experiments we will use RGB color channels and XY Spatial Location features. In superpixel Level Feature experiments we will use Mean of RGB color values, RGB color histogram, Mean of Gabor filter responses. You can reach implementation details and code on the project GitHub repository.

KMEANS CLUSTERING: K-means clustering is like sorting marbles into different groups based on their colors. Imagine you have a bunch of marbles in different shades of red, green, and blue, and you want to group them by their colors. K-means starts by randomly picking a few marbles as representatives, then assigns each marble to the closest representative based on color. After that, it recalculates the average color of the marbles in each group and moves the representative marble to that average color. It repeats this process, fine-tuning the groups until the marbles don't change groups much. Finally, you end up with clusters of marbles that are most similar in color within each group. In our experiment we have pixels instead of marbles. We will do same thing using different image features.

Discussion/Conclusion

As you can see on the result part we conducted many experiment. Be aware that this is image processing so results could be subjective but we want give brief explanation of result. We got best results on 3 and 4 clusters we couldn’t got good results and clear segments. As a method we got best results with histogram superpixels and Gabor filters of superpixels. When superpixel level is 25 we couldn’t got clear and viewable result so we can say that segment number should be greater than 50. Most of time when segment number is 50, 75 and 100 there isn’t much visually difference between results we can try increase n_neighbors on slice function.

Reference:

1. https://en.wikipedia.org/wiki/K-means_clustering
2. https://scikit- image.org/docs/stable/api/skimage.segmentation.html#skimage.segmentation.slic
3. https://people.uncw.edu/ricanekk/teaching/fall08/csc520/programming%20project.html

Source Code:

https://github.com/metehan41/IMAGE_ASS3.git