Graph-based image analysis

The objective of this work is to implement superpixel and Felzenszwalb-Huttenlocher clustering algorithms, push-relabel and Boykov-Kolmogorov segmenting algorithm and compare their performance with experiments. The first half of this work contains the explanation of used algorithms and how they work. Both clustering methods creates clusters made of pixels. There are three superpixel clustering algorithms: SLIC, SLICO, MSLIC and minimum spanning tree algorithm – Felzenszwalb-Huttenlocher. Both segmenting methods are searching for minimum cut in graph. Boykov-Kolmogorov creates search trees, while the push-relabel algorithm creates residual graph. In the second half of this work experiments were performed with the algorithms. The main point of the experiments was to observe and compare performance time and precision. From performance time experiments, push-relabel algorithm appears to be the one working faster. From precision experiments, the push-relabel algorithm has been noticed to give the best results, especially when contrast between object and background is small. During the experiments parameter values were found for best performance of the algorithms. To reduce Boykov-Kolmogorov algorithm performance time it is recommended to set the average superpixel time to 50 pixels, if better precision is wanted it is recommended to set it less than 50. For faster segmentation with push-relabel algorithm it is recommended to set minimum cluster size to 50 and scale to 100, if better precision is wanted, it is recommended to set it lower. Faster segmentation with these algorithms means lower precision and vice versa