This thesis addresses the problem of automatic detection and identification of blemishes in digital images of potatoes. Potatoes are an important food crop, with clear unblemished skin being the main factor affecting consumer preference. Potatoes with defects, diseases and blemishes caused by otherwise benign (to human) skin infections, are strongly avoided by consumers. Most potatoes are sorted into dfferent grades by hand, with inevitable mistakes and losses. The currently deployed computer vision systems for sorting potatoes require manual training and have limited accuracy and high unit costs. A further limitation of typical machine vision systems is that the set of image features for pattern recognition has to be designed by the system engineer to work with a specific configuration of produce, imaging system and operating conditions. Such systems typically do not generalise well to other configurations, where the required image features may well differ from those used to design the original system. The objective of the research presented in this thesis is to introduce an automatic method for detecting and identifying blemishes in digital images of potatoes, where the presented solution involves classifying individual pixels. A human expert is required to mark up areas of blemishes and non-blemishes in a set of training images. For blemish detection, each pixel is classified as either blemish or non-blemish. For blemish identification, each pixel is classified according to a number of pre-determined blemish categories. After training, the system should be able to classify individual pixels in new images of previously unseen potatoes with high accuracy. After segmenting the potato from the image background, a very large set of candidate features, based on statistical information relating to the colour and texture of the region surrounding a given pixel, is first extracted. The features include statistical summaries of the whole potato and local regions centred on each pixel as well as the data of the pixel itself. Then an adaptive boosting algorithm (AdaBoost) is used to automatically select the best features for discriminating between blemishes and non-blemishes. The AdaBoost algorithm (Freund and Schapire, 1999) is used to build a classifier, which combines results from so-called "weak" classifiers, each constructed using one of the candidate features, into one "strong" classifier that performs better than any of the weak classifiers alone. With this approach, different features can be selected for different potato varieties, while also handling the natural variation in fresh produce due to different seasons, lighting conditions, etc. For blemish detection, the classifier was trained using a subset of pixels which had been marked as blemish or non-blemish. Tests were done with the full set of features, "lesion experiments" were carried out to explore the impact of removing specific feature types, and experiments were also carried out on methods of speeding up classification both by restricting the number of weak classifiers and restricting the numbers of unique candidate features which can be used to produce weak classifiers. The results were highly accurate with visible examples of disagreement between classifier output and markup being caused by human inaccuracies in the markup rather than classifier inaccuracy. For blemish identification, a set of classifiers were trained on subsets of pixels marked as each blemish class against a subset of pixels drawn from all other classes. For classification, each pixel was tested with all classifiers and assigned to the classifier which returned the highest confidence of a positive result. Experiments were again performed with methods of speeding up classification as well as lesion experiments. Finally, to demonstrate how the system would work in an industrial context, the classification results were summarised for each potato, providing a high overall accuracy in detecting the presence or absence of significant blemish coverage for each blemish type.

Bird species are recognised as important biodiversity indicators: they are responsive to changes in sensitive ecosystems, whilst populations-level changes in behaviour are both visible and quantifiable. They are monitored by ecologists to determine factors causing population fluctuation and to help conserve and manage threatened and endangered species. Every five years, the health of bird population found in the UK are reviewed based on data collected from various surveys. Currently, techniques used in surveying species include manual counting, Bioacoustics and computer vision. The latter is still under development by researchers. Hitherto, no computer vision technique has fully been deployed in the field for counting species as these techniques use high-quality and detailed images of stationary birds, which make them impractical for deployment in the field, as most species in the field are in-flight and sometimes distant from the cameras field of view. Techniques such as manual and bioacoustics are the most frequently used but they can also become impractical, particularly when counting densely populated migratory species. Manual techniques are labour intensive whilst bioacoustics may be unusable when deployed for species that emit little or no sound. There is the need for automated systems for identifying species using computer vision and machine learning techniques, specifically for surveying densely populated migratory species. However, currently, most systems are not fully automated and use only appearance-based features for identification of species. Moreover, in the field, appearance-based features like colour may fade at a distance whilst motion-based features will remain discernible. Thus to achieve full automation, existing systems will have to combine both appearance and motion features. The aim of this thesis is to contribute to this problem by developing computer vision techniques which combine appearance and motion features to robustly classify species, whilst in flight. It is believed that once this is achieved, with additional development, it will be able to support the surveying of species and their behaviour studies. The first focus of this research was to refine appearance features previously used in other related works for use in automatic classification of species in flight. The bird appearances were described using a group of seven proposed appearance features, which have not previously been used for bird species classification. The proposed features improved the classification rate when compared to state-of-the-art systems that were based on appearance features alone (colour features). The second step was to extract motion features from videos of birds in flight, which were used for automatic classification. The motion of birds was described using a group of six features, which have not previously been used for bird species classification. The proposed motion features, when combined with the appearance features improved classification rates compared with only appearance or motion features. The classification rates were further improved using feature selection techniques. There was an increase of between 2-6% of correct classification rates across all classifiers, which may be attributable directly to the use of motion features. The only motion features selected are the wing beat frequency and vicinity features irrespective of the method used. This shows how important these groups of features were to species classification. Further analysis also revealed specific improvements in identifying species with similar visual appearance and that using the optimal motion features improve classification accuracy significantly. We attempt a further improvement in classification accuracy, using majority voting. This was used to aggregate classification results across a set of video sub-sequences, which improved classification rates considerably. The results using the combined features with majority voting outperform those without majority voting by 3% and 6% on the seven species and thirteen classes dataset respectively. Finally, a video dataset against which future work can be benchmarked has been collated. This data set enables the evaluation of work against a set of 13 species, enabling effective evaluation of automated species identification to date and a benchmark for further work in this area of research. The key contribution of this research is that a species classification system was developed, which combines motion and appearance features and evaluated it against existing appearance-only-based methods. This is not only the first work to combine features in this way but also the first to apply a voting technique to improve classification performance across an entire video sequence.

The plenoptic camera is an emerging technology in computer vision able to capture a light field image from a single exposure which allows a computational change of the perspective view just as the optical focus, known as refocusing. Until now there was no general method to pinpoint object planes that have been brought to focus or stereo baselines of perspective views posed by a plenoptic camera. Previous research has presented simplified ray models to prove the concept of refocusing and to enhance image and depth map qualities, but lacked promising distance estimates and an efficient refocusing hardware implementation. In this thesis, a pair of light rays is treated as a system of linear functions whose solution yields ray intersections indicating distances to refocused object planes or positions of virtual cameras that project perspective views. A refocusing image synthesis is derived from the proposed ray model and further developed to an array of switch-controlled semi-systolic FIR convolution filters. Their real-time performance is verified through simulation and implementation by means of an FPGA using VHDL programming. A series of experiments is carried out with different lenses and focus settings, where prediction results are compared with those of a real ray simulation tool and processed light field photographs for which a blur metric has been considered. Predictions accurately match measurements in light field photographs and signify deviations of less than 0.35 % in real ray simulation. A benchmark assessment of the proposed refocusing hardware implementation suggests a computation time speed-up of 99.91 % in comparison with a state-of-the-art technique. It is expected that this research supports in the prototyping stage of plenoptic cameras and microscopes as it helps specifying depth sampling planes, thus localising objects and provides a power-efficient refocusing hardware design for full-video applications as in broadcasting or motion picture arts.

Visual inspection of defects is an important part of quality assurance in many fields of production. It plays a very useful role in industrial applications in order to relieve human inspectors and improve the inspection accuracy and hence increasing productivity. Research has previously been done in defect classification of wood veneers using techniques such as neural networks, and a certain degree of success has been achieved. However, to improve results in tenus of both classification accuracy and running time are necessary if the techniques are to be widely adopted in industry, which has motivated this research. This research presents a method using rough sets based neural network with fuzzy input (RNNFI). Variable precision rough set (VPRS) method is proposed to remove redundant features utilising the characteristics of VPRS for data analysis and processing. The reduced data is fuzzified to represent the feature data in a more suitable foml for input to an improved BP neural network classifier. The improved BP neural network classifier is improved in three aspects: additional momentum, self-adaptive learning rates and dynamic error segmenting. Finally, to further consummate the classifier, a uniform design CUD) approach is introduced to optimise the key parameters because UD can generate a minimal set of uniform and representative design points scattered within the experiment domain. Optimal factor settings are achieved using a response surface (RSM) model and the nonlinear quadratic programming algorithm (NLPQL). Experiments have shown that the hybrid method is capable of classifying the defects of wood veneers with a fast convergence speed and high classification accuracy, comparing with other methods such as a neural network with fuzzy input and a rough sets based neural network. The research has demonstrated a methodology for visual inspection of defects, especially for situations where there is a large amount of data and a fast running speed is required. It is expected that this method can be applied to automatic visual inspection for production lines of other products such as ceramic tiles and strip steel.

Car make and model recognition (CMMR) has become an important part of intelligent transport systems. Information provided by CMMR can be utilized when licence plate numbers cannot be identified or fake number plates are used. CMMR can also be used when automatic identification of a certain model of a vehicle by camera is required. The majority of existing CMMR methods are designed to be used only in daytime when most car features can be easily seen. Few methods have been developed to cope with limited lighting conditions at night where many vehicle features cannot be detected. This work identifies car make and model at night by using available rear view features. A binary classifier ensemble is presented, designed to identify a particular car model of interest from other models. The combination of salient geographical and shape features of taillights and licence plates from the rear view are extracted and used in the recognition process. The majority vote of individual classifiers, support vector machine, decision tree, and k-nearest neighbours is applied to verify a target model in the classification process. The experiments on 100 car makes and models captured under limited lighting conditions at night against about 400 other car models show average high classification accuracy about 93%. The classification accuracy of the presented technique, 93%, is a bit lower than the daytime technique, as reported at 98 % tested on 21 CMMs (Zhang, 2013). However, with the limitation of car appearances at night, the classification accuracy of the car appearances gained from the technique used in this study is satisfied.

Diabetic retinopathy (DR) is a chronic, progressive and possibly vision-threatening eye disease. Early detection and diagnosis of DR, prior to the development of any lesions, is paramount for more efficiently dealing with it and managing its consequences. This thesis investigates and proposes a number of candidate geometric and haemodynamic biomarkers, derived from fundus images of the retinal vasculature, which can be reliably utilised for identifying the progression from diabetes to DR. Numerous studies exist in literature that investigate only some of these biomarkers in independent normal, diabetic and DR cohorts. However, none exist, to the best of my knowledge, that investigates more than 100 biomarkers altogether, both geometric and haemodynamic ones, for identifying the progression to DR, by also using a novel experimental design, where the same exact matched junctions and subjects are evaluated in a four year period that includes the last three years pre-DR (still diabetic eye) and the onset of DR (progressors’ group). Multiple additional conventional experimental designs, such as non-matched junctions, non-progressors’ group, and a combination of them are also adopted in order to present the superiority of this type of analysis for retinal features. Therefore, this thesis aims to present a complete framework and some novel knowledge, based on statistical analysis, feature selection processes and classification models, so as to provide robust, rigorous and meaningful statistical inferences, alongside efficient feature subsets that can identify the stages of the progression. In addition, a new and improved method for more accurately summarising the calibres of the retinal vessel trunks is also presented. The first original contribution of this thesis is that a series of haemodynamic features (blood flow rate, blood flow velocity, etc.), which are estimated from the retinal vascular geometry based on some boundary conditions, are applied to studying the progression from diabetes to DR. These features are found to undoubtedly contribute to the inferences and the understanding of the progression, yielding significant results, mainly for the venular network. The second major contribution is the proposed framework and the experimental design for more accurately and efficiently studying and quantifying the vascular alterations that occur during the progression to DR and that can be safely attributed only to this progression. The combination of the framework and the experimental design lead to more sound and concrete inferences, providing a set of features, such as the central retinal artery and vein equivalent, fractal dimension, blood flow rate, etc., that are indeed biomarkers of progression to DR. The third major contribution of this work is the new and improved method for more accurately summarising the calibre of an arterial or venular trunk, with a direct application to estimating the central retinal artery equivalent (CRAE), the central retinal vein equivalent (CRVE) and their quotient, the arteriovenous ratio (AVR). Finally, the improved method is shown to truly make a notable difference in the estimations, when compared to the established alternative method in literature, with an improvement between 0.24% and 0.49% in terms of the mean absolute percentage error and 0.013 in the area under the curve. I have demonstrated that some thoroughly planned experimental studies based on a comprehensive framework, which combines image processing algorithms, statistical and classification models, feature selection processes, and robust haemodynamic and geometric features, extracted from the retinal vasculature (as a whole and from specific areas of interest), provide altogether succinct evidence that the early detection of the progression from diabetes to DR can be indeed achieved. The performance that the eight different classification combinations achieved in terms of the area under the curve varied from 0.745 to 0.968.

Seismic layer segmentation, oil-gas boundary surfaces recognition, and 3D volume data reconstruction are three important tasks in three-dimensional seismic image processing. Geophysical and geological parameters and properties have been known to exhibit progressive changes in a layer. However, there are also times when sudden changes can occur between two layers. λ-connectedness was proposed to describe such a phenomenon. Based on graph theory, λ-connectedness describes the relationship among pixels in an image. It is proved that λ-connectedness is an equivalence relation. That is, it can be used to partition an image into different classes and hence can be used to perform image segmentation. Using the random graph theory and λ-connectivity of the image, the length of the path in a λ-connected set can be estimated. In addition to this, the normal λ-connected subsets preserve every path that is λ-connected in the subsets. An O(nlogn) time algorithm is designed for the normal λ-connected segmentation. Techniques developed are used to find objects in 2D/3D seismic images. Finding the interface between two layers or finding the boundary surfaces of an oil-gas reserve is often asked. This is equivalent to finding out whether a λ-connected set is an interface or surface. The problem that is raised is how to recognize a surface in digital spaces. λ-connectedness is a natural and intuitive way for describing digital surfaces and digital manifolds. Fast algorithms are designed to recognize whether an arbitrary set is a digital surface. Furthermore, the classification theorem of simple surface points is deduced: there are only six classes of simple surface points in 3D digital spaces. Our definition has been proved to be equivalent to Morgenthaler-Rosenfeld's definition of digital surfaces in direct adjacency. Reconstruction of a surface and data volume is important to the seismic data processing. Given a set of guiding pixels, the problem of generating a λ-connected (subset of image) surface is an inverted problem of λ-connected segmentation. In order to simplify the fitting algorithm, gradual variation, an equivalent concept of λ-connectedness, is used to preserve the continuity of the fitted surface. The key theorem, the necessary and sufficient condition for the gradually varied interpolation, has been mathematically proven. A random gradually varied surface fitting is designed, and other theoretical aspects are investigated. The concepts are used to successfully reconstruct 3D seismic real data volumes. This thesis proposes λ-connectedness and its applications as applied to seismic data processing. It is used for other problems such as ionogram scaling and object tracking. It has the potential to become a general technique in image processing and computer vision applications. Concepts and knowledge from several areas in mathematics such as Set Theory, Fuzzy Set Theory, Graph Theory, Numerical Analysis, Topology, Discrete Geometry, Computational Complexity, and Algorithm Design and Analysis have been applied to the work of this thesis.