A Study On Auto-thresholding Selection Methods For Image Segmentation

In the application of remote sensing images, people are often interested in specific information of the image. When extracting the specific target information, it is necessary to separate the target from the whole original image and try to avoid the disturbances from the background. How to rationally and effectively obtain the threshold which is used to differentiate the object from the background is the key process. However, thresholds based on different segmentation algorithms would affect the precision of the result and the detail level of the characteristic information. The traditional methods, such as human-computer interaction or using experienced values, are often influenced by subjective factors. They can neither determine the threshold automatically nor be applied to all situations. Therefore, studying on the automatic selection threshold methods is an important subject for image segmentation and worth further research.By now, there is no certain threshold algorithm suitable for all types of images. Certainly for specific images and specific requirements, we can find the optimum threshold methods with quite good effect. This paper studies and discusses several automatic selection threshold methods for simple and complex images separately. Since the latter is based on the former, this paper is focus on global threshold in simple images.Global threshold method based on pixel gray scale can be used in a simple remote sensing image. While iterative method can be used in the uniform illumination images which have double-peak and deep-valley histogram. When there are higher real-time and stability requirements, the minimum error method is a better choice. Maximum entropy method can not only be applied to the images with low gray contrast and single-peak histogram, but also be applied to small targets extraction; however, it is sensitivity to noise. For the images that with the suitable size of target, the maximum between-class variance method can be used to find the optimal threshold for image segmentation. Minimum cross-entropy threshold method focuses on minimizing the discrepancies of information between the original image and segmented image, it has high applicability in images with single-peak histogram as well as double-peak histogram; but, it is sensitivity to the size of the target.For the remote sensing images with much noise and non-uniform background characteristics, the threshold method based on attributes of pixel neighborhood is more appropriate. This article develops one-dimensional Otsu algorithm and one-dimensional maximum entropy method to two-dimensional space; and also proposes an algorithm based on edge weighting. All of them can effectively overcome the noise interference, and the segmentation results are better than the one-dimensional threshold algorithm. However, the computation time of the two-dimensional algorithm is much longer than the one-dimensional algorithm due to the multiplied computation.As for complex remote sensing images which have relatively rich gray levels, fuzzy boundaries, complex structure and different contrast, adaptive threshold method based on local properties can be used to segment the image. Mean method and Median method take the mean and median value of each pixel in an operator window as the threshold, which can greatly simplify the algorithm of traditional interpolation method. It takes the different characteristics of each part of the image into account, and can be success in segmentation.