Line Eikvil

Applications of hidden Markov chains in image analysis

In image analysis, two-dimensional Markov models, i.e. Markov field models, have been applied for segmentation purposes, but except for the area of text recognition, the application of hidden Markov chains has been rare. Through four very different examples, this paper demonstrates the applicability also for hidden Markov chains in image analysis, and shows that the problems of image analysis often may have one- dimensional characteristics even though the images are two-dimensional.

Tools for interactive map conversion and vectorization

The process of converting an analog map into structured digitized information requires several different operations, which are all time-consuming when performed manually. Strictly automatic processing is not always a possible solution, and an interactive approach can then be an alternative. The paper describes a tool for map conversion, focusing on the functionality for extraction of line structures. An interactive approach is used as it gives the user an opportunity to survey the process, and utilize human knowledge. The methods are based on contour following, extracting centre points needed for accurate vector representation of the line during tracing.

Statistical classification using a linear mixture of two multinormal probability densities

Abstract The paper describes an estimation-maximization algorithm to estimate the parameters of a probability density model consisting of a linear mixture of two multinormal distributions. Superior classification results to those obtained using the multinormal distribution or the k-nearest neighbour rule were obtained with this model on two difficult data sets.

Use of hidden Markov models and phenology for multitemporal satellite image classification: applications to mountain vegetation classification

Ground cover classification based on a single satel- lite image can be challenging. The work reported here concerns the use of multitemporal satellite image data in order to alleviate this problem. We consider the problem of vegetation mapping and model the phenological evolution of the vegetation using a Hidden Markov Model (HMM). The different vegetation classes can be in one of a predefined set of states related to their phenological development. The characteristics of a given class are specified by the state transition probabilities as well as the probability of given satellite observations for that class and state. Classification of a specific pixel is thus reduced to selecting the class that has the highest probability of producing a given series of observations for that pixel. Compared to standard classification techniques such as maximum likelihood (ML) classification, the proposed scheme is flexible in that it derives its properties not only from image specific training data, but also from a model of the temporal behavior of the ground cover. It is shown to produce results that compare favorably to those obtained using ML classification on single satellite images, it also generalizes better than this approach. Obtaining good ground cover classifications based on a single satellite image can be challenging. The work reported here concerns the use of multitemporal satellite image data in order to alleviate this problem. We will consider an application of these methods to mapping of high mountain vegetation in Norway. The traditional mapping method based on manual field work is prohibitively expensive and alternatives are therefore sought. Vegetation classification based on satellite images is an interesting alternative, but the complexity of the vegetation ground cover is high and the use of multitemporal satellite image acquisitions is shown to improve the classifi- cation quality. This document is organized as follows: In the next section, we briefly recapitulate previous work related to multitemporal satellite image classification and phenological models. In section IV we discuss the HMM and how it can be used for classification. In section V we show results of the application of our algorithm, conclusions are given in section VI.

Text page recognition using Grey-level features and hidden Markov models

This paper presents an approach to text recognition which avoids the problems of thresholding and segmentation by working directly on the grey-level image recognizing an entire word at the time. For each word a sequence of grey-level feature vectors is extracted. Hidden Markov models are used to describe the single characters and the sequence of feature vectors is matched against all possible combinations of models using dynamic programming.

Text Recogniton from Grey Level Images Using Hidden Markovc Models

The problems of character recognition are today mainly due to imperfect thresholding and segmentation. In this paper a new approach to text recognition is presented which attempts to avoid these problems by working directly on grey level images and treating an entire word at the time. The features are found from the grey levels of the image, and a hidden Markov model is defined for each character. During recognition the most probable combination of models is found for each word by the use of dynamic programming.

Adaptive Registration of Remote Sensing Images using Supervised Learning

This paper describes a system for co-registration of time series satellite images which uses a learning-based strategy. During a training phase, the system learns to recognize regions in an image suited for registration. It also learns the relationship between image characteristics and registration performance for a set of different registration algorithms. This enables intelligent selection of an appropriate registration algorithm for each region in the image, while regions unsuited for registration can be discarded. The approach is intended for co-registration of sequences of images acquired from identical or similar earth observation sensors. It has been tested for such sequences from different types of sensors, both optical and radar, with varying resolution. For images with moderate differences in content, the registration accuracy is, in general, good with an RMS error of one pixel or less.

Alignment of growth seasons from satellite data

This work concerns the alignment of growth seasons based on satellite data. This work is motivated by a high mountain vegetation classification problem in Norway. Vegetation classes are characterized by their temporal evolution through a growth season. Data of high spatial resolution, like LANDSAT data, are often temporally sparse. In order to get a longer sequence of images, data from different years can be combined into one single synthetic sequence. We describe a method for determining the correspondence between the chronological time of the image acquisition and the time at which the phenological state of the vegetation cover shown in the image would typically occur. The task is considered as a minimization problem and is solved by dynamic programming. The methodology is based on the normalized difference vegetation index (NDVI) computed from data having a coarse spatial resolution such as MODIS or AVHRR data. The proposed methodology has been tested on data from several years covering a region in Norway including mountainous areas. It is evident from plots of the original data that NDVI curves from different seasons are shifted relative to one another. By applying the proposed time warping methodology to adjust the time scale within each year the shifts become less apparent. We conclude that the methodology can be used for alignment of growth seasons from satellite data.

Evaluation of Binary Descriptors for Fast and Fully Automatic Identification

In this study we evaluate the potential of local binary descriptors for automatic sorting in an industrial context. This problem is different from that of retrieval for human handling as we need to identify the one correct class, rather than finding all the similar classes. We have looked at classes of objects that need to be identified by their cover or label, rather than their shape. Challenges for this application are that the process needs to be very fast and the approach must be able to distinguish between a large number of classes, where the classes can be quite similar and have identical elements. We have studied various combinations of detectors and binary descriptors in combination with approximate nearest neighbor (ANN) searches in such contexts. Our conclusion is that these approaches are well suited for this type of automatic sorting, and our experiments show that for the best performing combinations we are able to obtain a 99% recognition rate on a database of 80,000 images using an average of less than 0.5 seconds per image.

DECODING BAR CODES FROM HUMAN-READABLE CHARACTERS

Abstract In this study we have decoded bar codes by recognizing the human-readable characters of the interpretation line printed below the bar pattern. Using this approach, we were able to successfully decode bar codes with a resolution of 0.8 pixels per module.