Nicholas R. Howe

Document binarization with automatic parameter tuning

Document analysis systems often begin with binarization as a first processing stage. Although numerous techniques for binarization have been proposed, the results produced can vary in quality and often prove sensitive to the settings of one or more control parameters. This paper examines a promising approach to binarization based upon simple principles, and shows that its success depends most significantly upon the values of two key parameters. It further describes an automatic technique for setting these parameters in a manner that tunes them to the individual image, yielding a final binarization algorithm that can cut total error by one-third with respect to the baseline version. The results of this method advance the state of the art on recent benchmarks.

Silhouette Lookup for Automatic Pose Tracking

Computers should be able to detect and track the articulated 3-D pose of a human being moving through a video sequence. Current tracking methods often prove slow and unreliable, and many must be initialized by a human operator before they can track a sequence. This paper introduces a simple yet effective algorithm for tracking articulated pose, based upon looking up observed silhouettes in a collection of known poses. The new algorithm runs quickly, can initialize itself without human intervention, and can automatically recover from critical tracking errors made while tracking previous frames in a video sequence.

A Laplacian Energy for Document Binarization

This paper describes a new algorithm for document binarization, building upon recent work in energy-based segmentation methods. It uses the Laplacian operator to assess the local likelihood of foreground and background labels, Canny edge detection to identify likely discontinuities, and a graph cut implementation to efficiently find the minimum energy solution of an objective function combining these concepts. The results of this algorithm place it near the top on both the DIBCO-09 and H-DIBCO assessments.

Boosted decision trees for word recognition in handwritten document retrieval

Recognition and retrieval of historical handwritten material is an unsolved problem. We propose a novel approach to recognizing and retrieving handwritten manuscripts, based upon word image classification as a key step. Decision trees with normalized pixels as features form the basis of a highly accurate AdaBoost classifier, trained on a corpus of word images that have been resized and sampled at a pyramid of resolutions. To stem problems from the highly skewed distribution of class frequencies, word classes with very few training samples are augmented with stochastically altered versions of the originals. This increases recognition performance substantially. On a standard corpus of 20 pages of handwritten material from the George Washington collection the recognition performance shows a substantial improvement in performance over previous published results (75% vs 65%). Following word recognition, retrieval is done using a language model over the recognized words. Retrieval performance also shows substantially improved results over previously published results on this database. Recognition/retrieval results on a more challenging database of 100 pages from the George Washington collection are also presented.

Examining Locally Varying Weights for Nearest Neighbor Algorithms

Previous work on feature weighting for case-based learning algorithms has tended to use either global weights or weights that vary over extremely local regions of the case space. This paper examines the use of coarsely local weighting schemes, where feature weights are allowed to vary but are identical for groups or clusters of cases. We present a new technique, called class distribution weighting (CDW), that allows weights to vary at the class level. We further extend CDW into a family of related techniques that exhibit varying degrees of locality, from global to local. The class distribution techniques are then applied to a set of eleven concept learning tasks. We find that one or more of the CDW variants significantly improves classification accuracy for nine of the eleven tasks. In addition, we find that the relative importance of classes, features, and feature values in a particular domain determines which variant is most successful.

Silhouette lookup for monocular 3D pose tracking

Computers should be able to detect and track the articulated 3D pose of a human being moving through a video sequence. Incremental tracking methods often prove slow and unreliable, and many must be initialized by a human operator before they can track a sequence. This paper describes a simple yet effective algorithm for tracking articulated pose, based upon looking up observations (such as body silhouettes) within a collection of known poses. The new algorithm runs quickly, can initialize itself without human intervention, and can automatically recover from critical tracking errors made while tracking previous frames in a video sequence.

Part-Structured Inkball Models for One-Shot Handwritten Word Spotting

Many document collections of historical interest are handwritten and lack transcripts. Scholars need tools for high-quality information retrieval in such environments, preferably without the burden of extensive system training. This paper presents a novel approach to word spotting designed for manuscripts or degraded print that requires minimal initial training. It can infer a generative word appearance model from a single instance, and then use the model to retrieve similar words from arbitrary documents. An approximation to the retrieval statistic runs efficiently on graphics processing hardware. Tested on two standard data sets, the method compares favorably with prior results.

Finding words in alphabet soup: Inference on freeform character recognition for historical scripts

This paper develops word recognition methods for historical handwritten cursive and printed documents. It employs a powerful segmentation-free letter detection method based upon joint boosting with histograms of gradients as features. Efficient inference on an ensemble of hidden Markov models can select the most probable sequence of candidate character detections to recognize complete words in ambiguous handwritten text, drawing on character n-gram and physical separation models. Experiments with two corpora of handwritten historic documents show that this approach recognizes known words more accurately than previous efforts, and can also recognize out-of-vocabulary words.

Percentile blobs for image similarity

We present a new algorithm called PBSIM for computing image similarity, based upon a novel method of extracting bloblike features from images. In tests on a classification task using a data set of over 1000 images, PBSIM shows significantly higher accuracy than algorithms based upon color histograms, as well as previously reported results for another approach based upon bloblike features.

A closer look at boosted image retrieval

Margin-maximizing techniques such as boosting have been generating excitement in machine learning circles for several years now. Although these techniques offer significant improvements over previous methods on classification tasks, little research has examined the application of techniques such as boosting to the problem of retrieval from image and video databases. This paper looks at boosting for image retrieval and classification, with a comparative evaluation of several top algorithms combined in two different ways with boosting. The results show that boosting improves retrieval precision and recall (as expected), but that variations in the way boosting is applied can significantly affect the degree of improvement observed. An analysis suggests guidelines for the best way to apply boosting for retrieval with a given image representation.