Clustering is one of the image-processing methods used in non-destructive testing (NDT). As one of the initializing parameters, most clustering algorithms, like fuzzy C means (FCM), Iterative self-organization data analysis (ISODATA), K-means, and their derivatives, require the number of clusters. This paper proposes an algorithm for clustering the pixels in C-scan images without any initializing parameters. In this state-of-the-art method, an image is sampled based on the rosette pattern and according to the pattern characteristics, and extracted samples are clustered and then the number of clusters is determined. The centroids of the classes are computed by means of a method used to calculate the distribution function. Based on different data sets, the results show that the algorithm improves the clustering capability by 92.93% and 91.93% in comparison with FCM and K-means algorithms, respectively. Moreover, when dealing with high-resolution data sets, the efficiency of the algorithm in terms of cluster detection and run time improves considerably.