Shaun Timothy Love

Printer banding estimation using-the Generalized Spectrum

This paper compares multi-step algorithms for estimating banding arameters of a harmonic signature model. The algorithms are based on two different spectral measures, the power spectrum (PS) and the collapsed average (CA) of the generalized spectrum. The generalized spectrum has superior noise reduction properties and is applied for the first time to this application. Monte Carlo simulations compare estimation performances of profile (or coherent) averaging and non-coherent spatial averaging for estimating banding parameters in grain noise. Results demonstrate that profile averaging has superior noise reduction properties, but is less flexible in applications with irregular banding patterns. The PS-based methods result in lower fundamental frequency estimation error and greater peak height stability for low SNR values, with coherent averaging being significantly superior to non-coherent averaging. The CA has the potential of simplifying the detection of multiple simultaneous banding patterns because its peaks are related to intra-harmonic distances; however, good CA estimation performance requires sufficiently regular harmonic phase patterns for the banding harmonics so as not to undergo reduction along with the noise. In addition to the simulations, the algorithms are applied to samples from inkjet and laser printers to demonstrate the ability of the harmonic signature model in separating banding from grain and other image artifacts. Good results from experimental data are demonstrated based on visual inspection of examples where banding and grain have been separated.

Adaptive image interpolation using a multilayer neural network

Image resizing is an important operation that is used extensively in document processing to magnify or reduce images. Standard approaches fit the original data with a continuous model and then resample this 2D function on a few sampling grid. These interpolation methods, however, apply an interpolation function indiscriminately to the whole image. The resulting document image suffers from objectionable moire patterns, edge blurring and aliasing. Therefore, image documents must often be segmented before other document processing techniques, such as filtering, resizing, and compression can be applied. In this paper, we present a new system to segment and label document images into text, halftone images, and background using feature extraction and unsupervised clustering. Once the segmentation is performed, a specific enhancement or interpolation kernel can be applied to each document component. In this paper, we demonstrate the power of our approach to segment document images into text, halftone, and background. The proposed filtering and interpolation method results in a noticeable improvement in the enhanced and resized image.

Document image segmentation using a two-stage neural network

In this paper, we present a new system to segment and label document images into text, halftone images, and background using feature extraction and unsupervised clustering. Each pixel is assigned a feature pattern consisting of a scaled family of differential geometrical invariant features and texture features extracted from the cooccurence matrix. The invariant feature pattern is then assigned to a specific region using a two-stage neural network system. The first stage is a self-organizing principal components analysis (SOPCA) network that is used to project the feature vector onto its leading principal axes found by using principal components analysis. Using the SOPCA algorithm, we can train the SOPCA network to project our feature vector orthogonally onto the subspace spanned by the eigenvectors belonging to the largest eigenvalues. By doing that we ensure that the vector is represented by a reduced number of effective features. The next step is to cluster the output of the SOPCA network into different regions. This is accomplished using a self-organizing feature-map (SOFM) network. In this paper, we demonstrate the power of the SOPCA-SOFM approach to segment document images into text, halftone, and background.

An image quality motivated complex content classifier

Image quality loss is often determined by the nature and level of image artifacts along with the image context they appear in. For example, grain may be masked by texture, and blur is tolerable in flat fields, but offensive in regions of edges and structure. This paper develops image region classifiers for complex (real life) images. Based on the contents structure, the classes of interest are: a random field (such as sky or painted surfaces); textured regions (such as grass or line textures); regions with transients (such as edges on buildings). The linear classifiers examined use features from the optical density histogram (ODH), the cortex transform, and the co-occurrence matrix. The performance testing of the classifiers show that the best feature set size is four. Larger sets show no classification error reduction and tend to suffer from overfitting. The best performance is 3.3% misclassification, and is achieved using four features from the ODH and cortex transform. A misclassification rate of 10% is achieved using only co-occurrence matrix features. This rate drops to 4.4%, when ODH, cortex transform, and co-occurrence features are combined. The classifiers were trained on image regions assigned to each of the three classes by human observers, then tested on a larger non-overlapping image region set.

Converting color values using stochastic interpolation

This paper presents a method of efficiently converting from a set of noisy color values to a set of device colorants. Using a deterministic process, 24-bit scanned color values are reduced to dithered 12-bit RGB table indices. After the reduction, a small but complete lookup table with 4096 entries converts the RGB values directly to the output color space. This stochastic interpolation process, while minimizing banding and abrupt color transitions, eliminates the need for trilinear interpolation of the data and significantly reduces the size of the lookup table.