Fig.l Various regions of a document Document decomposition is a process where various regions of any scanned image or scanned text document are identified and categorized. Its main aim is to distinguish the textual zone from non-textual zone. Documents’ layout may vary depending on the sources (newspaper, text books, magazines etc.) they have been taken from. The key parts of any document are identified through document decomposition. Like for a newspaper key parts are images, headlines, columns, sub-headlines. For document decomposition, at first we analysed the layout of taken input. We all know that each page maintains a certain gap between the margins and the textual zone. We used this characteristic of pages to find out the area from where the main document is started. Then we detected the edges of each element. Next, At first, we took an input image in form of matrix. Then we found out the edges in the input image and mark them in the output map using the ‘Canny’ algorithm. Before applying this algorithm, it was ensured that, the image and the edges of various portions of the image remains noiseless and fairly sharp respectively. Fig.2 shows the input image on left side and the edge detected image on right side. In this step, we pick every pixel of the input and assigned them a value according to the length of the vertical and horizontal line it belongs to. That is every pixel of same line will be given a same value. Here we assigned 0 (black) for pixel to the array if the pixel color is not black (>0). Thus we will get an image where, the textual area will be black and the not textual area will be white. The converted image for the Fig.2 is shown in Fig.3. Fig.3 Converted image We proposed three filters to define a region as image. First filter checks, whether the length and width of each block is greater than a threshold value or not. Second one finds the pixel ratio of the block and compares them with a threshold value. Third filter determines the ratio of the sum of the total length of each block and total width of each block. The detected images of Fig 2 are shown here. length of each block and total width of each block. The detected images of Fig 2 are shown here. 3.5.2 Labelling the Text Zones width of each block is greater than a threshold value or not. Second one finds the pixel ratio of the To categorise the regions of text zones into headline, sub headline and columns we moved ahead with the methodology of dominant line/letter height. We checked out the line heights of inputs from several ‘epapers’. We found out the gap between height of dominant letters, height of headlines, and sub headlines. To detect the letter heights perfectly, we removed the image portion from the input after detection. Through those results, we decided some threshold value ‘gap1’, ‘gap2’, ‘x1’, ‘x2’, ‘x3’ to label the text regions. To categorise the regions of text zones into headline, sub headline and columns we moved ahead We proposed three filters to define a region as image. First filter checks, whether the length and Fig.7 Skewed and rotated input, de-skewed and de-rotated output 4. RESULT ANALYSIS We checked about 70 skewed images. The accuracy of our de- skew system is approximately 100%. W checked 70 skewed as well as rotated images. Then the accuracy was 97.83%. We decomposed about 300 scanned newspaper pages. The precision, accuracy and recall of our methodolog are given below. Input may be both skewed and rotated. To solve this problem, we computed the skew angle of the input. Input may contain an image portion at the beginning. For such cases to compute the skew angle perfectly the input was rotated to 90°, 180° and -90°. Then skew angles of all these rotated inputs were calculated. We picked the highest absolute value of skew angles and de-skewed the input according to this value. After this the de- skewed image was auto rotated. For this we calculated the line height, matra line, index point, total black pixel on matra line for both the input image and 90° rotated image. Then we compared the data computed from previous step. We compared the whether ‘the half of the line heig' we decided whether we have to pixel ratio (pixel/line height), and took the bigger value. Now, we checked ht’ is greater than ‘the matra line index’ or not. In addition, depending on this rotate the image or not. If rotation is needed, we will return 180° or -90°. depending on which pixel ratio value we chose. Table 1 Result Analysis Table 5. CONCLUSION
