Clustering segmentation method realizes image segmentation by clustering analysis of image gray features or other high-order statistical features, which can be divided into two categories: hard clustering and soft clustering. Hard clustering represented by K-means considers that each pixel only belongs to a certain class, and calculates the distance between the pixel and each cluster center at runtime, and if the distance from the center of a certain class is the smallest, the pixel Classify into this category, and then iteratively optimize until the mean value of each cluster does not change. The soft clustering algorithm represented by fuzzy C-means considers that the image is inaccurate, incomplete, and fuzzy, and considers that each pixel belongs to multiple categories to a certain extent, so it calculates the number of pixels belonging to each category. Possibility, according to which the pixel is judged as the most likely category. In comparison, the clustering effect of the soft clustering algorithm is better than that of the hard clustering algorithm, but the amount of calculation is significantly increased.

 

The biggest difficulty of the cluster segmentation method is that it is difficult to determine the exact number of clusters. If the set number of clusters is too small, a large number of background areas in the image will be misclassified as foreground areas, and if the number is too large, it will be easily affected by noise. In addition, the initial value of the clustering center also has an important impact on the clustering results, if it is not set properly, it may lead to clustering failure. Therefore, clustering algorithms often need prior knowledge about images in practice, so as to add constraints in feature space analysis.

 

When applying the cluster segmentation method to segment thermal infrared images, due to the extremely unbalanced distribution of the histogram of the infrared image, the distribution of the target and the background in the histogram is not a simple bimodal shape in most cases, and it is often difficult to obtain a reasonable value by simply using the clustering algorithm. Segmentation Results]. For this reason, adopting improved clustering algorithms such as K-means clustering result correction can help to improve the effect of clustering segmentation.