Research On Image Thresholding Algorithms Based On Otsu Method

The aim of the image segmentation is to extract the interested object from the complex background for the scene analysis and object recognition. In the image processing area, thresholding is an age-old and also a classic basic problem, but is still a hotspot. Existing thresholding segmentation algorithms can be divided into global and local thresholding segmentation algorithms. When the object and background in an image with a histogram of bimodal distribution, global thresholding segmentation algorithms have a good case. This type of methods are simple and the real-time is very high, but because they are processing for the whole image, the segmented results are poor usually due to the very uneven illumination and other factors. Thus, global methods can not adapt to local changes. Typical global algorithms such Otsu method, minimum error method, maximum entropy method, et al. Local thresholding segmentation algorithms, namely adaptive thresholding algorithms, are based on pixels or regions of an image for processing. This type of methods are able to adapt to local image changes, they are of robustness. But the processing time for local approaches is longer, and need to artificially set the parameters. Classical local algorithms such as Bernsen method. Niblack method, Sauvola method, Gato method, and so on.In our work, some research of related algorithms is proceeding based on Otsu method. After analysing the advantage and disadvantage of the existing thresholding approaches, some improved schemes are presented. This paper has made the following key findings:(1) For the existing thresholding algorithms, they do not make full use of advantage of the mean, median and gradient information of an image, so we propose a mean-median-gradient three-dimensional co-occurrence matrix model. Then a Otsu thresholding method is introduced based on this model combining the Otsu approach. This 3-D model uses the mean and median filter, thus effectively reducing the interference such as Gaussian noise, salt and pepper noise or they mixed noise for the segmented image. The introduction of gradient information compensates the useful information which are filtered out after the corresponding mean, median filtering, making the segmentation of images retain more detailed information. Finally, a large number of experiments compared with gray-gradient method, mean-gradient the maximum correlation method,2-D minimum error method.2-D Renyi entropy.2-D Otsu method and 3-D Otsu method, the segmentation results show that our method have more robust segmentation performance.(2) In order to improve the poor anti-noise of the Otsu thresholding algorithm, the researchers introduced two-dimensional and three-dimensional approaches, and a set of improvement and fast algorithms respectively. Although the anti-noise performance has been improved with increasing of the dimension to some extent, the thresholding results are still poor. The real-time of this methods is very poor. A novel Otsu thresholding algorithm based on rebuilding and dimension reduction of the three-dimensional was proposed with the consideration of the poor anti-noise performance in 3-D Otsu's method due to the obviously wrong region division. Firstly, our method rebuilt the 3-D histogram based on the distribution of noise points, which reduced the noise interference. And then, we transfer the region division in 3-D histogram from eight partitions into two partitions, thus reducing the search space of threshold from 3-dimension to 1-dimension, which saved a lot of processing time and memory space. Finally, segmentation result and running time were given at the end of this paper. In addition, comparations with those of 3-D Otsu's method,2-D decomposing approach and 2-D oblique algorithm were given as well. The results show that the presented method has better anti-noise performance and visual quality compared with the above three approaches, and lower time complexity compared with 3-D Otsu's method.(3) Taking into account the uneven effects of light on image segmentation, compared to the global threshold segmentation algorithms, the local threshold segmentation algorithms have a more satisfactory segmentation results. However, due to the current adaptive threshold segmentation algorithms are sensitive to parameters which are usually acquired through experience. However, it is difficult to get to the best parameters, and for different images, the best parameters are completely different, so this largely limited the adaptive threshold segmentation algorithm processing performance. Considering the interactive segmentation idea can reduce the problem of parameters selection, an interactive Otsu Thresholding algorithm is presented based on image regions with the combination of interactive segmentation ideas, which the global thresholding algorithm is applied to the local area. The proposed method divides the image into several regions roughly and sorts all the image blocks according to their deviation values. And then the users input interactive information to separate all regions into three parts. The last step is to complete the binarization of every block. Extensive experiments are performed and the results show that the proposed scheme not only has high performance but also yields promising binarization outcomes for images under normal and inadequate illumination conditions, and reduces the sensitivity for parameters.