An Improved ${k}$-Means Algorithm with Spatial Constraints for Image Segmentation
Meng Hu, Eric C.C. Tsang, Yanting Guo, Qingshuo Zhang · 2021
The${k}$-means is sensitive to the initial center and noise in image segmentation. To reduce the sensitivity of k-means to the initial center, we use the equidistant strategy to segment the cumulative sum of histogram to initialize centers. Then we introduce local spatial information into objective function of k-means to reduce the impact of noise in image segmentation and improve the robustness of segmentation algorithm. We combine the equidistant segmentation strategy and k-means with spatial constraints to propose an improved k-means algorithm (I-${k}$-means_S). I-${k}$-means_S solves the problem that traditional${k}$-means is easy to be affected by initial center and noise in image segmentation. To test the performance of I-${k}$-means, we create two synthetic images and add Gaussian and Salt & Pepper noises to the two images. We compare I-${k}$-means_S with 5 classical cluster algorithms on the two noisy images. From the results, we know that the proposed algorithm is more robust than the other algorithms. Meanwhile, we apply I-${k}$-means_S to natural image processing. The results show that it still has high robustness.