| Summary: | 碩士 === 淡江大學 === 資訊工程學系 === 90 === This thesis studies and discusses the binary image recovery method of history documents. Although the processing time of using global thresholding method to apply image segmentation technique when processing historical document recovery will be faster and this method can erase noise under the condition of usual noise efficiently. But the performance of global thresholding method will degrade down and be inefficiently when the noise is too high. Whereas the letter’s border will be detected by using edge-based segmentation method, but a lot of dots of noise border will be presented at the same time. Nevertheless, connecting intermittent borders and letting the enclosed boundary of the connected borders to appear in the maximum gradient location will be a great challenge. However, we can use region-based segmentation method to erase the noise efficiently, but the border of detected words will lead to the unusual phenomena and the size of extracted words will be vary from original words.
We propose, in this thesis, a so-called Watershed-aided Image Segmentation (WIS) method to improve above-mentioned disadvantages efficiently. The steps of image segmentation in proposed WIS method can be presented in four parts: The first step uses the so-called Watershed Method to segment the image on the position with the local maximum variant intensity. The second step, in order to reduce the interference of noise, WIS depends on the average value of gray level to smooth the small regions which is segmented by Watershed Method. The third step gathers statistics of gray value of all pixels except the pixels on the watershed line in the image then smoothes histogram by the distribution of histogram and selects the global threshold as a valley value between the first two peaks automatically to threshold regions of the image. The last step classifies the pixels on the watershed line according to global threshold.
This thesis proposed a watershed-aided image segmentation method that can erase the interference of noise efficiently and the border of extracted words just locate in the maximum gradient value by comparing with other segmentation methods. Furthermore, WIS finds out the edge of words accurately and the size of these words fits the original word size closely.
|
|---|
