Research On Segmentation And Skew Detection Of Multi-region Document Images

In the modern information society, computer technology has been involved in various fields of our lives. The Internet has also become popular increasingly, and we depend on computers to get information more than ever before, a lot of work is shifted on to computer. Studying how to covert the traditional paper into electronic text has become a topic of concern. In daily life, there are a large number of documents to be handled. All of these documents include not only text files but also images and mixed files, so how to put them into computer efficiently and accurately has become urgent requirements.The main purpose of this thesis is to study algorithms for page segmentation and skew detection of multi-region document images. The thesis summarizes the common algorithms of page segmentation, and gives their advantages and disadvantages of each algorithm. Generally, methods of page segmentation can be classified into two types, one is structural analysis, and the other is texture analysis. The structural analysis includes top-down, bottom-up and a mixing of the two. The thesis presents two top-down methods, run-length smoothing and projection profile cut, and two bottom-up methods, neighborhood line density and connected component analysis. In addition, it gives several algorithms which usually be used in image segmentation.According to these algorithms, this paper presents an improved method of the projection profile cut algorithm. This algorithm solves the problem that the projection profile cut algorithm couldn't deal with complicated documents containing skewed multi-regions. First, the image is binarized, then denoised by erosion and dilation operation of mathematical morphology. Applying the improved projection profile cut algorithm to document images, we can find the cut-off points of the image which don't have any peak-valley point on the X-axis and Y-axis. With these cut-off points we could cut the image into small pieces, and then we conduct the same operation until multi-regions are separated.Skew estimating methods can be classified into five general categories: Hough transform, cross-correlation, projection profile, Fourier transform and nearest-neighbor, of which Fourier transform is rarely used because of its high complexity.During document scanning, the image may lose something inevitably, and the edges are not smoothing. If we use the normal image edge detection to find the profile, it increases not only the amount of computation but also many unnecessary calculations. The thesis proposed a brief method to find the profile of the image, for which there is no need to find the edges accurately, just to find the area which contains the image. The area being found is called bounding box. The thesis used GA algorithm to detect skew angles of the images. This method uses the area of the bounding box as its fitness function, in which only the coordinate values of the 4 corners need to be found. This can reduce tremendous computing complexity. Experimental results show that the proposed algorithm can certainly guarantee the accuracy for document image deskewing.