| Resumen |
Image segmentation is the fundamental basis of image analysis, in which a digital image is partitioned (i.e. segmented) into distinct sets of pixels that share local characteristics, such as intensity, color, or texture. The techniques for performing such partitioning can be classified into threshold-based, edge-based, and region-based segmentation methods. There are some region-based hybrid techniques that allow multilevel segmentation instead of binary segmentation at the cost of exponentially increasing their computational complexity as the Mean Shift segmentation technique. Also, determining appropriate values for their parameters can be non-trivial, so they have limitations. In this work, we propose a novel multilevel image segmentation algorithm that does not rely on a fixed predefined number of segments but rather enables the definition of the desired level of detail by means of a single parameter called the spatial variation range. Our algorithm first partitions its input image into a grid with N cells; then, it analyzes each cell using a metaheuristic method based on the Particle Swarm Optimization (PSO) algorithm, which does not search for the best global solution. Instead, each swarm (one for each grid cell) moves over the two-dimensional pixel space in the image, searching for high-density colors that may be representative of the largest possible pixel neighborhood within the cell boundaries. Then, the colors found are clustered and simplified, both at the neighborhood and cell levels. Finally, the results obtained by all swarms are simplified globally, and the resulting color set is used to replace the color of each pixel in the original image with its closest match in the set, thus generating the segmented image. The proposed algorithm was tested on some images of the COCO (Common Object in Context) data set. Our preliminary results show that this method effectively identifies the main objects and enhances their uniformity within the images, achieving processing times equal to or better than the Fast Fuzzy C-Means Clustering for Image Segmentation algorithm (SFFCM) |
