Document page classification and nonlinear diffusion filtering for image segmentation and noise removal
We propose a novel family of nonlinear diffusions and apply it tothe problem of segmentation of multivalued images defined on arbitrary graphs. These equations generalize previously proposed scalar-valued diffusion equations which have been used for segmenting grayscale images. We demonstrate the effectiveness of our new methods on a large number of color, texture and natural image segmentation tasks. We in addition introduce another extension of our diffusion equations which can process orientation images, i.e., images whose every pixel takes values on a circle. We analyze a specific scalar-valued diffusion equation and show that it can be used as an approximate method for solving the constrained total variation minimization problem posed by Osher, Rudin, and Fatemi. We illustrate the resulting algorithm by applying it to noise removal problems. We develop a real-time, strip-based, low-complexity document page classification algorithm, which can be used as a copy mode selector in the copy pipeline. It analyzes the scanned RGB images and classifies them into one of eight modes. Modes are the combinations of mono/color and text/mix/picture/photo settings. Mode classification is 35% to 99% accurate with misclassifications tending towards benign modes. The benefits of such a copy mode selector include improving copy quality, simplifying user interaction, and increasing copy rate. We also propose an alternative algorithm working on the image data represented in an opponent color space. The algorithm reduces the computational complexity by a factor of three, and maintains very close classification results to the algorithm designed for the $RGB$ color space.
Pollak, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our