Image Analysis and Recognition

In plant development biology, the study of the structure of the plant’s root is fundamental for the understanding of the regulation and interrelationships of cell division and cellular differentiation. This is based on the high connection between cell length and progression of cell differentiation and the nuclear state. However, the need to analyse a large amount of images from many replicate roots to obtain reliable measurements motivates the development of automatic tools for root structure analysis.

We present a novel automatic approach to detect cell files, the main structure in plant roots, and extract the length of the cells in those files. This approach is based on the detection of local cell file characteristic symmetry using a wavelet based image symmetry measure.

The resulting detection enables the automatic extraction of important data on the plant development stage and of characteristics for individual cells. Furthermore, the approach presented reduces in more than 90% the time required for the analysis of each root, improving the work of the biologist and allowing the increase of the amount of data to be analysed for each experimental condition.

While our approach is fully automatic a user verification and editing stage is provided so that any existing errors may be corrected. Given five test images it was observed that user did not correct more than 20% of all automatically detected structure, while taking no more than 10% of manual analysis time to do so.

To perform high-throughput single-cell studies, automation of the potential experiments is quite necessary. Due to their complex morphology, automatic manipulation and visual analysis of adherent cells which include a wide range of mammalian cell lines is a challenging task. In this paper, the problem of adherent cells localization for the purpose of automated robotic microinjection has been stated and a practical two-stage texture-based solution has been proposed. The method has been tested on NIH/3T3 cells and the results have been reported.

Producing consistent segmentations of lung nodules in CT scans is a persistent problem of image processing algorithms. Many hard-segmentation approaches are proposed in the literature, but soft segmentation of lung nodules remains largely unexplored. In this paper, we propose a classification-based approach based on pixel-level texture features that produces soft (probabilistic) segmentations. We tested this classifier on the publicly available Lung Image Database Consortium (LIDC) dataset. We further refined the classification results with a post-processing algorithm based on the variability index. The algorithm performed well on nodules not adjacent to the chest wall, producing a soft overlap between radiologists’ based segmentation and computer-based segmentation of 0.52. In the long term, these soft segmentations will be useful for representing the uncertainty in nodule boundaries that is manifest in radiological image segmentations.

Although segmentation of biomedical image data has been paid a lot of attention for many years, this crucial task still meets the problem of the correctness of the obtained results. Especially in the case of optical microscopy, the ground truth (GT), which is a very important tool for the validation of image processing algorithms, is not available.

We have developed a toolkit that generates fully 3D digital phantoms, that represent the structure of the studied biological objects. While former papers concentrated on the modelling of isolated cells (such as blood cells), this work focuses on a representative of tissue image type, namely human colon tissue. This phantom image can be submitted to the engine that simulates the image acquisition process. Such synthetic image can be further processed, e.g. deconvolved or segmented. The results can be compared with the GT derived from the digital phantom and the quality of the applied algorithm can be measured.

Diabetes mellitus is a disease that may cause dysfunctions in the sympathetic and parasympathetic nervous system. Therefore, the pupillary reflex of diabetic patients shows characteristics that distinguish them from healthy people, such as pupil radius and contraction time. These features can be measured by the noninvasive way of dynamic pupillometry, and an analysis of the data can be used to check the existence of a neuropathy. In this paper, it is proposed the use of artificial neural networks for helping screening the diabetes occurrence through the dynamic characteristics of the pupil, with successful results.

Active contour, due to acceptable results in the field of image segmentation, has attracted more attention in the last several decades. However, the low quality and the presence of noise in medical images, particularly ultrasound images have also created some limitations for this method, such as Entrapment within the local minima and adjustment of the contour coefficients. In this paper, we present a segmental algorithm combined active contour and genetic algorithm to remove these limitations and bring some improvements to the segmentation outcome. The experimental results show that our proposed algorithm has an acceptable accuracy.

We propose a technical solution that enables 3D video-based in-bore movement quantification to be acquired synchronously with the BOLD function magnetic resonance imaging (fMRI) sequences. Our solution relies on in-bore video setup with 2 cameras mounted in a 90 degrees angle that allows tracking movments while acquiring fMRI sequences. In this study we show that using 3D motion quantification of a simple finger opposition paradigm we were able to map two different finger positions to two different BOLD response patterns in a typical block design protocol. The motion information was also used to adjust the block design to the actual motion start and stop improving the time accuracy of the analysis. These results reinforce the role of video based motion quantification in fMRI analysis as an independent regressor that allows new findings not discernable when using traditional block designs.

This paper proposes a classification-based method for automating the segmentation of the region of interest (ROI) in digital images of chromatographic plates. Image segmentation is performed in two phases. In the first phase an unsupervised learning method classifies the image pixels into three classes: frame, ROI or unknown. In the second phase, distance features calculated for the members of the three classes are used for deciding on the new label, ROI or frame, for each individual connected segment previously classified as unknown.The segmentation result is post-processed using a sequence of morphological operators before obtaining the final ROI rectangular area. The proposed methodology, which is the initial step for the development of a screening tool for Fabry disease, was successfully evaluated in a dataset of 58 chromatographic images.

This paper presents a novel method for the automatic pupil and iris localization. The proposed algorithm is based on an automatic adaptive thresholding method that iteratively looks for a region that has the highest chances of enclosing the pupil. Once the pupil is localized, next step is to find the boundary of iris based on the first derivative of each row of the area within the pupil. We have tested our proposed algorithm on two public databases namely: CASIA v1.0 and MMU v1.0 and experimental results show that the proposed method has satisfying performance and good robustness against the reflection in the pupil.

This paper addresses the problem of local ridge orientation estimation of fingerprint image. The proposed method computes multiple candidate orientations for each foreground block. A systematic method, consisting of orientation voting and orientation propagation, for selecting orientation out of the multiple candidates is designed according to the smooth changing of local ridge orientations. Experiments show that the proposed methods are robust for poor quality images and the overall matching performance is improved.

This paper discusses the advantages of score-level fusion between pattern and minutiae based fingerprint verification algorithms in various operational scenarios. The different scenarios considered are sensor interoperability, environmental conditions and low quality enrollments. These are commonly encountered in real-life deployments of fingerprint-based biometric systems, specifically for large-scale distributed systems and physical access control. Moreover, the approach for jointly utilizing the conceptually different pattern and minutiae algorithms is based on various well-known scorelevel fusion techniques with single finger presentations. In contrast to previous studies on multi-matcher score-level fusion for fingerprint verification, where only moderate performance improvement were reported, the results presented here show significant performance gains. The two main contributing factors to these findings are that the two algorithms are conceptually different and the effects of the different operational scenarios. For the latter, improvement in accuracy due to fusion is even more significant in non-ideal and challenging operating conditions.

This paper presents an improved image-based fingerprint verification system. The proposed system enhances an input fingerprint image using a contextual filtering technique in the frequency domain, and uses the complex fillers to identify the core point. Subsequently, a region of interest (ROI) of a predefined size, which is centered around the detected core point, is extracted. The resulting ROI is rotated based on the detected core point angle to ensure rotation invariance. The proposed system extracts the absolute average deviation from the outputs of eight oriented Gabor filters that are applied to the ROI. To reduce the dimensionality of the extracted features whilst generating more discriminatory representation, this paper compares the unsupervised principal component analysis and the supervised linear discriminant analysis methods for dimensionality reduction. User-specific thresholding schemes are investigated. The effectiveness of the proposed algorithm is evaluated on the public FVC2002 set_a database. Experimental results demonstrate the superiority of the introduced solution in comparison with existing approaches.

Gender classification based on facial images has received increased attention in the computer vision literature. Previous work on this topic has focused on images acquired in the visible spectrum (VIS). We explore the possibility of predicting gender from face images acquired in the near-infrared spectrum (NIR). In this regard, we address the following two questions: (a) Can gender be predicted from NIR face images; and (b) Can a gender predictor learned using VIS images operate successfully on NIR images and vice-versa? Our experimental results suggest that NIR face images do have some discriminatory information pertaining to gender, although the degree of discrimination is noticeably lower than that of VIS images. Further, the use of an illumination normalization routine may be essential for facilitating cross-spectral gender prediction.

The paper considers new approach to palm shape analysis that is based on continuous skeletons of binary images. The approach includes polygonal approximation of binary image, skeleton construction for the polygonal approximation and skeleton regularization by pruning. Skeleton of polygonal shape is a locus of centers of maximum inscribed circles. Both internal and external skeletons of palm shape are used for analysis. Segmentation of initial image, palm orientation and structure identification, fingers segmentation and characteristic points detection are performed based on image skeleton. Algorithm of color palm images binarization and computational experiments with large database of such images are described in the paper.

Effective analysis and recognition of human emotional behavior are important for achieving efficient and intelligent human computer interaction. This paper presents an approach for audiovisual based multimodal emotion recognition. The proposed solution integrates the audio and visual information by fusing the kernel matrices of respective channels through algebraic operations, followed by dimensionality reduction techniques to map the original disparate features to a nonlinearly transformed joint subspace. A hidden Markov model is employed for characterizing the statistical dependence across successive frames, and identifying the inherent temporal structure of the features. We examine the kernel fusion method at both feature and score levels. The effectiveness of the proposed method is demonstrated through extensive experimentation.

Several computational systems which depend on the precise location of the eyes have been developed in the last decades. Aware of this need, we propose a method for automatic detection of eyes in images of human faces using four geostatistical functions - semivariogram, semimadogram, covariogram and correlogram and support vector machines. The method was tested using the ORL human face database, which contains 400 images grouped in 40 persons, each having 10 different expressions. The detection obtained the following results of sensibility of 93.3%, specificity of 82.2% and accuracy of 89.4%.

Any information about people such as their gender may be useful in some secure places; however, in some occasions, it is more appropriate to obtain such information in an unobtrusive manner such as using gait. In this study, we propose a novel method for gender classification using gait template, which is based on Radon Transform of Mean Gait Energy Image (RTMGEI). Robustness against image noises and reducing data dimensionality can be achieved by using Radon Transformation, as well as capturing variations of Mean Gait Energy Images (MGEIs) over their centers. Feature extraction is done by applying Zernike moments to RTMGEIs. Orthogonal property of Zernike moment basis functions guarantee the statistically independence of coefficients in extracted feature vectors. The obtained feature vectors are used to train a Support Vector Machine (SVM). Our method is evaluated on the CASIA database. The maximum Correct Classification Rate (CCR) of 98.94% was achieved for gender classification. Results show that our method outperforms the recently presented works due to its high performance.

In this paper, we present a person re-identification method based on appearance classification. It consists a human silhouette comparison by characterizing and classification of a persons appearance (the front and the back appearance) using the geometric distance between the detected head of person and the camera. The combination of head detector with an orthogonal iteration algorithm to help head pose estimation and appearance classification is the novelty of our work. In this way, the is achieved robustness against viewpoint, illumination and clothes appearance changes. Our approach uses matching of interest-points descriptors based on fast cross-bin metric. The approach applies to situations where the number of people varies continuously, considering multiple images for each individual.

Matching Short Wave InfraRed (SWIR) face images against a face gallery of color images is a very challenging task. The photometric properties of images in these two spectral bands are highly distinct. This work presents a new cross-spectral face recognition method that encodes both magnitude and phase of responses of a classic bank of Gabor filters applied to multi-spectral face images. Three local operators: Simplified Weber Local Descriptor, Local Binary Pattern, and Generalized Local Binary Pattern are involved. The comparison of encoded face images is performed using the symmetric Kullbuck-Leibler divergence. We show that the proposed method provides high recognition rates at different spectra (visible, Near InfraRed and SWIR). In terms of recognition rates it outperforms Faceit®G8, a commercial software distributed by L1.

Face recognition in real-world applications is often hindered by uncontrolled settings including pose, expression, and illumination changes, and/or ageing. Additional challenges related to changes in facial appearance due to plastic surgery have become apparent recently. We exploit the fact that plastic surgery bears on appearance in a non-uniform fashion using a recognition approach that integrates information derived from local region analysis. We implemented and evaluated the performance of two new integrative methods, FARO and FACE, which are based on fractals and a localized version of a correlation index, respectively Experimental results confirm the expectation that face recognition is indeed challenged by the effects of plastic surgery. The same experimental results also show that both FARO and FACE compare favourably against standard face recognition methods such as PCA and LDA.

Sparse representation for face recognition has been exploited in past years. Several sparse representation algorithms have been developed. In this paper, a novel eyeglasses-face recognition approach, SEMD Based Sparse Gabor Representation, is proposed. Firstly, for a robust representation to misalignment, a sparse Gabor representation is proposed. Secondly, spatially constrained earth mover’s distance is employed instead of Euclidean distance to measure the similarity between original data and reconstructed data. The proposed algorithm for eyeglasses-face recognition has been evaluated under different eyeglasses-face databases. The experimental results reveal that the proposed approach is validity and has better recognition performance than that obtained using other traditional methods.

Face images obtained by an outdoor surveillance camera, are often confronted with severe degradations (e.g., low-resolution, low-contrast, blur and noise). This significantly limits the performance of face recognition (FR) systems. This paper presents a framework to overcome the degradation in images obtained by an outdoor surveillance camera, to improve the performance of FR. We have defined a measure that is based on the difference in intensity histograms of face images, to estimate the amount of degradation. In the past, super-resolution techniques have been proposed to increase the image resolution for face recognition. In this work, we attempt a combination of partial restoration (using super-resolution, interpolation etc.) of probe samples (long distance shots of outdoor) and simulated degradation of gallery samples (indoor shots). Due to the unavailability of any benchmark face database with gallery and probe images, we have built our own database and conducted experiments on a realistic surveillance face database.

PCA

and

FLDA

have been used as baseline face recognition classifiers. The aim is to illustrate the effectiveness of our proposed method of compensating the degradation in surveillance data, rather than designing a specific classifier space suited for degraded test probes. The efficiency of the method is shown by improvement in the face classification accuracy, while comparing results obtained separately using training with acquired indoor gallery samples and then testing with the outdoor probes.

A novel image-level fusion algorithm is proposed for 3D face recognition, which synthesizes an integrate image from both 2D intensity and 3D depth images. Due to the same descriptors in 2D and 3D domain, the image combination not only maintains facial intrinsic details to the utmost extent, but also provides more distinctive features. Also as the result of image recognition, the low efficiency of 3D surface matching is eliminated, and a fast 3D face recognition system is carried out. After the proposed surface preprocessing, an enhanced ULBP descriptor is applied to reduce the feature dimension, and LDA is adopted to extract the optimal discriminative components from the integrate image. Experiments performed on the FRGC v2.0 show that this algorithm practically outperforms the existing state-of-art multimodel recognition algorithm and realizes a real-time face recognition system.

Automatic facial expression analysis is the most commonly studied aspect of behavior understanding and human-computer interface. Most facial expression recognition systems are implemented with general expression models. However, the same facial expression may vary differently across humans, this can be true even for the same person when the expression is displayed in different contexts. These factors present a significant challenge for recognition. To cope with this problem, we present in this paper a personalized facial action recognition framework that we wish to use in a clinical setting with familiar faces; in this case a high accuracy level is required. The graph fitting method that we are using offers a constrained tracking approach on both shape (using procrustes transformation) and appearance (using weighted Gabor wavelet similarity measure). The tracking process is based on a modified Gabor-phase based disparity estimation technique.

Experimental results show that the facial feature points can be tracked with sufficient precision leading to a high facial expression recognition performance.

In this paper we present a new prediction technique to compress a pair of satellite images that have significant overlap in the underlying spatial areas. When this prediction technique is combined with an existing lossless image set compression algorithm, the results are significantly better than those obtained by compressing each image individually. Even when there are significant differences between the two images due to factors such as seasonal and atmospheric variations, the new prediction technique still performs very well to achieve significant reduction in storage requirements.

In this paper, a Vector Quantization compression scheme based on block indexing is proposed to compress true color images. This scheme uses affine transform to represent the blocks of the image in terms of the blocks of the code book. In this work a template image rich with high contrast areas is used as a codebook to approximately represent the blocks of the compressed image. A time reduction was achieved due to the usage of block descriptors to index the images blocks, these block descriptors are derived from the discrete cosine transform (DCT) coefficients. The DCT bases descriptor is affine transform invariant. This descriptor is used to filter out the domain blocks, and make matching only with similar indexed blocks. This introduced method led to time (1.13sec), PSNR (30.09), MSE (63.6) and compression ratio (7.31) for Lena image (256×256, 24bits).

Numerical experiments indicate that images, in general, possess a considerable degree of affine self-similarity, that is, blocks are well approximated in root mean square error (RMSE) by a number of other blocks when affine greyscale transformations are employed. This has led to a simple

L

2

-based model of affine image self-similarity which includes the method of fractal image coding (cross-scale, affine greyscale similarity) and the nonlocal means denoising method (same-scale, translational similarity). We revisit this model in terms of the structural similarity (SSIM) image quality measure, first deriving the optimal affine coefficients for SSIM-based approximations, and then applying them to various test images. We show that the SSIM-based model of self-similarity removes the “unfair advantage” of low-variance blocks exhibited in

L

2

-based approximations. We also demonstrate experimentally that the local variance is the principal factor for self-similarity in natural images both in RMSE and in SSIM-based models.

A fair P2P scalable video streaming scheme is proposed in this paper. The contributions of the paper are threefold. First, to improve the quality fairness of multiple video streams, a modified Lloyd-Max quantization-based Priority Index (PID) assignment method for scalable video coding (SVC) is developed, where the base-layer quality is also embedded in the Supplemental Enhancement Information (SEI) message of the SVC bit stream, in addition to the quantized rate-distortion slope of each MGS packet. Secondly, to build a fair ”contribute-and-reward” mechanism for P2P video streaming, we propose a multi-hierarchical topology that is based on peers’ uploading bandwidth. Finally, we combine these two parts to build a SVC-based P2P network, which fully utilizes the quality scalability of SVC, and the end-user quality is determined by its uploading bandwidth contribution. The performance of the scheme is demonstrated by simulation results.

In this paper, a novel yet simple encryption technique is proposed based on toral automorphism, Markov map and singular value decomposition (SVD). The core idea of the proposed scheme is to scramble the pixel positions by the means of toral automorphism and then encrypting the scrambled image using Markov map and SVD. The combination of Markov map and SVD changed the pixels values significantly in order to confuse the relationship among the pixels. Finally, a reliable decryption scheme is proposed to construct original image from encrypted image. Experimental results demonstrate the efficiency and robustness of the proposed scheme.

The paper presents an improved method for estimating the number of moving people in a scene for video surveillance applications; the performance is measured on the public database used in the framework of the PETS international competition, and compared, on the same database, with the ones participating to the same contest up to now. The system exhibits a high accuracy, ranking it at the top positions, and revealed to be so fast to make possible its use in real time surveillance applications.

In this paper, a land-cover extraction thematic mapping approach for urban areas from very high resolution aerial images is presented. Recent developments in the field of sensor technology have increased the challenges of interpreting images contents particularly in the case of complex scenes of dense urban areas. The major objective of this study is to improve the quality of land-cover classification. We investigated the use of multiple classifier systems (MCS) based on dynamic classifier selection. The selection scheme consists of an ensemble of weak classifiers, a trainable selector, and a combiner. We also investigated the effect of using Particle Swarm Optimization (PSO) based classifier as the base classifier in the ensemble module, for the classification of such complex problems. A PSO-based classifier discovers the classification rules by simulating the social behaviour of animals. We experimented with the parallel ensemble architecture wherein the feature space is divided randomly among the ensemble and the selector. We report the results of using separate/similar training sets for the ensemble and the selector, and how each case affects the global classification error. The results show that selection improves the combination performance compared to the combination of all classifiers with a higher improvement when using different training set scenarios and also shows the potential of the PSO-based approach for classifying such images.

Heat scintillation occurs due to the index of refraction of air decreasing with an increase in air temperature, causing objects to appear blurred and waver slowly in a quasi-periodic fashion. This imposes limitations on sensors used to record images over long distances resulting in a loss of detail in the video sequences. A method of filtering turbulent sequences using grid smoothing is presented that can be used to either extract a single geometrically improved frame or filter an entire turbulent sequence. The extracted frame is in general sharper than when utilising simple GFATR (Generalized First Average Then Register). It also better preserves edges and lines as well as being geometrically improved.

In this paper, a new image based method for detecting and extracting events in noisy hydrophone data sequence is developed. The method relies on dominant orientation and its robust reconstruction based on mutual information (MI) measure. This new reconstructed dominant orientation map of the spectrogram image can provide key segments corresponding to various acoustic events and is robust to noise. The proposed method is useful for long-term monitoring and a proper interpretation for a wide variety of marine mammals and human related activities using hydrophone data. The experimental results demonstrate that this image based approach can efficiently detect and extract unusual events, such as whale calls from the highly noisy hydrophone recordings.

This paper presents an improved method for detecting and segmenting taillight pairs of multiple preceding cars in busy traffic in day as well as night. Novelties and advantages of this method are that it is designed to detect multiple car simultaneously, it does not require knowledge of lanes, it works in busy traffic in daylight as well as night, and it is fast irrespective of number of preceding vehicles in the scene, and therefore suitable for real-time applications. The time to process the scene is independent of the size of the vehicle in pixels, and the number of preceding cars detected.

One of the previous night taillight detection methods in literature is modified to detect taillight pairs in the scene for both day and night conditions. This paper further introduces a novel hypothesis verification method based on the mathematical relationship between the vehicle distance from the vanishing point and the location of and distance between its taillights. This method enables the detection of multiple preceding vehicles in multiple lanes in a busy traffic environment in real-time. The results are compared with state-of-the-art algorithms for preceding vehicle detection performance, time and ease of implementation.

This study deals with the estimation of the road surface markings and their class using an onboard camera in an Advanced Driver Assistance System (ADAS). The proposed classification is performed in 3 successive steps corresponding to 3 levels of abstraction from the pixel to the object level through the connected-component one. At each level, a Markov Random Field models the a priori knowledge about object intrinsic features and object interactions, in particular spatial interactions. The proposed algorithm has been applied to simulated data simulated in various road configurations: dashed or continuous lane edges, road input, etc. These first results are very promising.

The huge amount of information some of the new optical satellites developed nowadays will create demands to quickly and reliably compensate for changes in the atmospheric transmittance and varying solar illumination conditions. In this paper three different forms of affine transformation models (general, particular and diagonal) are considered as candidates for rapid compensation of illumination variations. They are tested on a group of three pairs of CHRIS-PROBA radiance images obtained in a test field in Barrax (Spain), and where there is a difference in the atmospheric as well as in the geometrical acquisition conditions. Results indicate that the proposed methodology is satisfactory for practical normalization of varying illumination and atmospheric conditions in remotely sensed images required for operational applications.

The crevasse, which has always been one of the most dangerous factors on the Antarctic continent, threatens the life of the team members during the polar expedition. Crevasse detection is thus an increasingly important issue as it facilitates the analysis of glaciers and ice cap movements, research on the effects of climate change, and improves security for expedition staff. In this paper, we first analyze the characteristics of crevasse in ASTER image. We then test five features: Gray-Level Co-occurrence Matrices (GLCM), Gabor filters, Local Phase Quantization (LPQ), the completed local binary pattern (CLBP), and local self-similarity (LSS) for crevasse detection with the SVM classifier. Finally, we evaluate and validate the detection performance on two datasets. Experimental results show that the LSS descriptor performs better than other descriptors, and is thus a promising feature descriptor for crevasse detection.

In the paper the problem of the recognition of trademarks placed on banners which are visible during a sport television broadcast is described and experimentally investigated. It constitutes the second stage of the process of the analysis of the banners, e.g. in order to estimate the time that a particular banner is visible and can exert influence on the customers’ behaviour. Banners placed near a play field during football matches were analysed. For this task four algorithms were selected and tested, namely the UNL shape descriptor combined with the Partial Point Matching Algorithm, the Contour Sequence Moments, the UNL-Fourier descriptor and the Point Distance Histogram. Amongst them the best result was obtained when using the UNL + PPMA approach. The average efficiency of this method was equal to 84%.

In order to successfuly navigate between rows of plants, automated strawberry harvesters require a robust and accurate method of measuring the distance between the harvester and the strawberry bed. A diffracted red laser is used to project a straight horizontal line onto the bed, and is viewed by a video camera positioned at an angle to the laser. Using filtering techniques and the Hough transform, the distance to the bed can be calculated accurately at many points simultaneously, allowing the harverster’s navigation system to determine both its position and angle relative to the bed. Testing has shown that this low-cost solution provides near-perfect field performance.

Arabic handwritten text recognition has not received the same attention as that directed towards Latin script-based languages. In this paper, we present our efforts to develop a comprehensive Arabic Handwritten Text database (AHTD). At this stage, the database will consist of text written by 1000 writers from different countries. Currently, it has data from over 300 writers. It is composed of an images database containing images of the written text at various resolutions, and a ground truth database that contains meta-data describing the written text at the page, paragraph, and line levels. Tools to extract paragraphs from pages, segment paragraphs into lines have also been developed. Segmentation of lines into words will follow. The database will be made freely available to researchers world-wide. It is hoped that the AHTD database will stir research efforts in various handwritten-related problems such as text recognition, and writer identification and verification.