Converting handwritten or machine printed documents into a computer readable format allows more efficient storage and processing. The recognition of machine printed text is very reliable with today's technology, but the recognition of offline handwriting still remains a problem to the research community due to the high variance in handwriting styles. Modern mobile devices are capable of performing complex tasks such as scanning invoices, reading traffic signs, and online handwriting recognition, but there are only a few applications that treat offline handwriting. This thesis investigates the segmentation of handwritten documents into text lines and words, how the legibility of handwriting can be increased by beautification, as well as implementing it for modern mobile devices. Text line and word segmentation are crucial steps towards implementing a complete handwriting recognition system. The results of this thesis show that text line and word segmentation along with handwriting beautification can be implemented successfully for modern mobile devices and a survey concluding that the writing on processed documents is more legible than their unprocessed counterparts. An application for the operating system iOS is developed for demonstration.

This thesis focuses on improving the user experience for computer vision-based Augmented Reality (AR) applications on smartphones. The first part shows our proposed methods to enhance image binarisation. This improves the marker detection results in mobile AR applications. The comparisons of the original ARToolKit binarization method, our proposed histogram-based automatic thresholding and our histogram equalization based thresholding show that the histogram-based automatic thresholding produces a relatively better result under extreme and normal lighting conditions but slightly reduces the ARToolKit framerate. The second part introduces a new fast painterly rendering algorithm which produces an immersive experience for mobile AR users. The proposed algorithm has low complexity and achieves a real-time performance on smartphones. In addition, this study has carried out a preliminary experiment comparing mobile GPU-based image processing algorithms with CPU-based equivalent on smartphones. The result indicates that the GPU-based implementations perform better than the CPU when processing relatively large sized images.

Outdoor tracking and registration are important enabling technologies for mobile augmented reality. Sensor fusion and image processing can be used to improve global tracking and registration for low-cost mobile devices with limited computational power and sensor accuracy. Prior research has confirmed the benefits of this approach with high-end hardware, however the methods previously used are not ideal for current consumer mobile devices. We discuss the development of a hybrid tracking and registration algorithm that combines multiple sensors and image processing to improve on existing work in both performance and accuracy. As part of this, we developed the Transform Flow toolkit, which is one of the first open source systems for developing and quantifiably evaluating mobile AR tracking algorithms. We used this system to compare our proposed hybrid tracking algorithm with a purely sensor based approach, and to perform a user study to analyse the effects of improved precision on real world tracking tasks. Our results show that our implementation is an improvement over a purely sensor fusion based approach; accuracy is improved up to 25x in some cases with only 2-4ms additional processing per frame, in comparison with other algorithms which can take over 300ms.

Boundary detection is an essential first-step for many computer vision applications. In practice, boundary detection is difficult because most images contain texture. Normally, texture-boundary detectors are complex, and so cannot run in real-time. On the other hand, the few texture boundary detectors that do run in real-time leave much to be desired in terms of quality. This thesis proposes two real-time texture-boundary detectors – the Variance Ridge Detector and the Texton Ridge Detector – both of which can detect high-quality texture-boundaries in real-time. The Variance Ridge Detector is able to run at 47 frames per second on 320 by 240 images, while scoring an F-measure of 0.62 (out of a theoretical maximum of 0.79) on the Berkeley segmentation dataset. The Texton Ridge Detector runs at 10 frames per second but produces slightly better results, with an F-measure score of 0.63. These objective measurements show that the two proposed texture-boundary detectors outperform all other texture-boundary detectors on either quality or speed. As boundary detection is so widely-used, this development could induce improvements to many real-time computer vision applications.

The maximum subarray problem (MSP) involves selection of a segment of consecutive array elements that has the largest possible sum over all other segments in a given array. The efficient algorithms for the MSP and related problems are expected to contribute to various applications in genomic sequence analysis, data mining or in computer vision etc. The MSP is a conceptually simple problem, and several linear time optimal algorithms for 1D version of the problem are already known. For 2D version, the currently known upper bounds are cubic or near-cubic time. For the wider applications, it would be interesting if multiple maximum subarrays are computed instead of just one, which motivates the work in the first half of the thesis. The generalized problem of K-maximum subarray involves finding K segments of the largest sum in sorted order. Two subcategories of the problem can be defined, which are K-overlapping maximum subarray problem (K-OMSP), and K-disjoint maximum subarray problem (K-DMSP). Studies on the K-OMSP have not been undertaken previously, hence the thesis explores various techniques to speed up the computation, and several new algorithms. The first algorithm for the 1D problem is of O(Kn) time, and increasingly efficient algorithms of O(K² + n logK) time, O((n+K) logK) time and O(n+K logmin(K, n)) time are presented. Considerations on extending these results to higher dimensions are made, which contributes to establishing O(n³) time for 2D version of the problem where K is bounded by a certain range. Ruzzo and Tompa studied the problem of all maximal scoring subsequences, whose definition is almost identical to that of the K-DMSP with a few subtle differences. Despite slight differences, their linear time algorithm is readily capable of computing the 1D K-DMSP, but it is not easily extended to higher dimensions. This observation motivates a new algorithm based on the tournament data structure, which is of O(n+K logmin(K, n)) worst-case time. The extended version of the new algorithm is capable of processing a 2D problem in O(n³ + min(K, n) · n² logmin(K, n)) time, that is O(n³) for K ≤ n/log n For the 2D MSP, the cubic time sequential computation is still expensive for practical purposes considering potential applications in computer vision and data mining. The second half of the thesis investigates a speed-up option through parallel computation. Previous parallel algorithms for the 2D MSP have huge demand for hardware resources, or their target parallel computation models are in the realm of pure theoretics. A nice compromise between speed and cost can be realized through utilizing a mesh topology. Two mesh algorithms for the 2D MSP with O(n) running time that require a network of size O(n²) are designed and analyzed, and various techniques are considered to maximize the practicality to their full potential.

Biometric identification is a pattern recognition based classification system that recognizes an individual by determining its authenticity using a specific physiological or behavioural characteristic (biometric). In contrast to number of commercially available biometric systems for human recognition in the market today, there is no such a biometric system for plant recognition, even though they have many characteristics that are uniquely identifiable at a species level. The goal of the study was to develop a plant species biometric using both global and local features of leaf images. In recent years, various approaches have been proposed for characterizing leaf images. Most of them were based on a global representation of leaf peripheral with Fourier descriptors, polygonal approximations and centroid-contour distance curve. Global representation of leaf shapes does not provide enough information to characterise species uniquely since different species of plants have similar leaf shapes. Others were based on leaf vein extraction using intensity histograms and trained artificial neural network classifiers. Leaf venation extraction is not always possible since it is not always visible in photographic images. This study proposed a novel approach of leaf identification based on feature hierarchies. First, leaves were sorted by their overall shape using shape signatures. Then this sorted list was pruned based on global and local shape descriptors. The consequent biometric was tested using a corpus of 200 leaves from 40 common New Zealand broadleaf plant species which encompass all categories of local information of leaf peripherals. Two novel shape signatures (full-width to length ratio distribution and half-width to length ratio distribution) were proposed and biometric vectors were constructed using both novel shape signatures, complex-coordinates and centroid-distance for comparison. Retrievals were compared and the biometric vector based on full-width to length ratio distribution was found to be the best classifier. Three types of local information of the leaf peripheral (leaf margin coarseness, stem length to blade length ratio and leaf tip curvature) and the global shape descriptor, leaf compactness, were used to prune the list further. The proposed biometric was able to successfully identify the correct species for 37 test images (out of 40). The proposed biometric identified all the test images (100%) correctly if two species were returned compared to the low recall rates of Wang et al. (2003) (30%, if 10 images were returned) and Ye et al. (2004) (71.4%, if top 5 images were returned). The biometric can be strengthened by adding reference images of new species to the database, or by adding more reference images of existing species when the reference images are not enough to cover the leaf shapes.

Breast cancer is one of the most prevalent forms of cancer in the world today. The search for effective treatment and screening methods is a highly active area of research. The Digital Image-based ElastoTomography (DIET) project is a new breast cancer screening system under development, where surface motion from the mechanically actuated breast is measured in 3D, and used as input to an inverse problem solving for breast elasticity. Cancerous lesions appear as high contrast features, being an order of magnitude stiffer than healthy tissue. The 3D motion capture is measured by an array of digital cameras using computer vision techniques. This thesis presents a complete imaging system and algorithms for the capture of 3D breast surface motion. The main components of the 3D motion capture system are the hardware and software image capture system, camera calibration, intra-image feature tracking, and 3D surface and motion reconstruction. Accurate algorithms for each of these components are developed, with a view to future development and potential modifications needed for a clinically-appropriate system. A number of the algorithms developed have potential applications outside of the DIET system. Proof of concept studies demonstrate the viability of the system, with full motion reconstruction being performed on silicone gel phantoms, designed to approximate human soft tissue, in a number of laboratory experiments.

The smart bookshelf serves as a test-bed to study environments that are intelligently augmented by projector-camera devices. The system utilizes a camera pair and a projector coupled with an RFID reader to monitor and maintain the state of a real world library shelf. Using a simple calibration scheme, the homography induced by the world plane in which book spines approximately lie is estimated. As books are added to the shelf, a foreground detection algorithm which takes into account the projected information yields new pixels in each view that are then verified using a planar parallax constraint across both cameras to yield the precise location of the book spine. The system allows users to query for the presence of a books through a user interface, highlighting the spines of present book using the known locations obtained through foreground detection and transforming image pixels to their corresponding points in the projectors frame via a derived homography. The system also can display the state of the bookshelf at any time in the past. Utilizing RFID tags increases robustness and usefulness of the application. Tags encode information about a book such as the title, author, etc, that can be used to query the system. It is used in conjunction with the visual system to infer the state of the shelf. This work provides a novel foreground detection algorithm that works across views, using loose geometric constraints instead pixel color similarity to robustly isolate foreground pixels. The system also takes into account projected information which if not handled would be detrimental to the system. The intent of this work was to study the feasibility of an augmented reality system and use this application as a testbed to study the issues of building such a system.

We explore the integration of projected imagery with a physical book that acts as a tangible interface to multimedia data. Using a camera and projector pair, a tracking framework is presented wherein the 3D position of planar pages are monitored as they are turned back and forth by a user, and data is correctly warped and projected onto each page at interactive rates to provide the user with an intuitive mixed-reality experience. The book pages are blank, so traditional camera-based approaches to tracking physical features on the display surface do not apply. Instead, in each frame, feature points are independently extracted from the camera and projector images, and matched to recover the geometry of the pages in motion. The book can be loaded with multimedia content, including images and videos. In addition, volumetric datasets can be explored by removing a page from the book and using it as a tool to navigate through a virtual 3D volume.

Carotenoids are essential for visual function, however their potential role in Computer Vision Syndrome (CVS) is not known. By providing carrot powder, this study examined carotenoid metabolism and visual function in CVS. CVS participants were recruited into a double-blind, placebo-controlled, repeated measures trial (n=19, ages 20-65) and were randomized to 2 supplementation groups; control (15g cream of wheat powder) or carrot enriched (15g carrot powder, 33% of vitamin A RDA for adults) in an isocaloric pudding and yogurt for 4 weeks. Retinal function, self-perceived vision status, and plasma carotenoids/retinoids were assessed, along with plasma lipids and oxidative stress markers. Photopic b-waves marginally improved following supplementation reflecting higher phototransduction, possibly due to increased plasma carotenoid/retinoid levels. LDL cholesterol and oxidative stress markers showed trending reductions illustrating a protective role of the carrot. Carrot powder, at a minimal supplementation dose, can be recommended for CVS.