Image thresholding based on semivariance

In this paper, an algorithm for image thresholding based on semivariance analysis is presented. The rationale of the approach is to binarize an image such that it best reproduces the original image variation across several spatial scales. The method can be alternatively viewed as one identifying the binary image that best approximate the overall level of edgeness measured across multiple scales in the original image. A comparison with seven other thresholding methods is presented for 2 synthetic images and 22 Non-Destructive Testing (NDT) grey level images. The results indicate that the proposed method is highly competitive. Performance of the proposed method in relation to the image content is also discussed.

► An algorithm for image thresholding based on semivariance analysis is presented.
► Threshold selected so that bilevel variogram best fits that of original image.
► The method is evaluated on 2 synthetic and 19 real-world images.
► Results demonstrate effectiveness of the approach, as compared with 4 other methods.