This Bachelor thesis aims at implementing the watershed algorithm to achieve the required segmentation of an image. The implementation is done both sequentially and in a parallelized approach using the NVIDIA CUDA architecture. A review of the chosen architecture is presented exploring its important aspects. Image segmentation and the most famous algorithms of implementing it are explained. A deeper look is taken on the watershed algorithm and its different approaches. A previously proposed algorithm [VKALF10] for the implementation of the watershed-based image segmentation is used. All the implementation steps are explained in details focusing on the parallelized way of their implementation. Several optimization techniques were implemented and tested to achieve the highest performance possible. Speedups reaching up to 15 times faster parallel GPU execution over the sequential CPU execution were achieved. The results obtained are examined and used to further analyze the difference between the parallel implementation and the sequential one. The effect of the proposed optimizations were analyzed and examined as well.