Combined method for scanned documents images segmentation using sequential extraction of regions

We propose a combined method to segment the images of scanned documents, which, in contrast to known methods, implies a preliminary separation of the graphics and photograph regions from the text regions and a background. In this case, an analysis of the connected components is performed, which are different for graph­ics, photographs, and text regions. In order to classify the selected regions into the photograph and graphics regions, a block method is employed. It was established that such a technique for splitting the regions into blocks less affects the quality of segmentation when compared to applying the block method directly to the original im­age. To extract the text regions that are more complex in their shape from the background, the neighborhood of each pixel was processed. To detect the boundaries of illustrations on the images of scanned documents, we applied the bloomberg method. In order to classify into photographs and graphics, it is proposed to split an illustration into blocks of pixels. Each block of pixels is identified with a vector of two features: the mean value of the local gradient magnitude, and the mean value of the function that localizes at the images of scanned documents the linear objects (graphics and text characters). The derived feature vectors were classified using a sup­port vector machine. When extracting the text regions, we applied a low-frequency filtering and a thresholding. The combined method was implemented in practice to segment the test images of scanned newspaper articles from the document da­tabase mediateam at oulu university (finland). It was established that the combined method is characterized by an increase in perfor­mance speed during image segmentation at high quality processing.