Dans cette thèse, on introduit la métrique "Coefficient de forme" pour la classement des données de dissimilitudes. Cette approche est inspirée par l'analyse discriminante géométrique et on a défini des règles de décision pour imiter le comportement du classifieur linéaire et quadratique. Le nombre de paramètres est limité (deux par classe). On a également étendu et amélioré cette démarche avantageuse et rapide pour apprendre uniquement à partir des représentations de dissimilitudes en utilisant l'efficacité du classificateur des Machines à Vecteurs de Support. Comme contexte applicatif pour la classification par dissimilitudes, on utilise la recherche d'images à l'aide d'une représentation des images en multi échelle en utilisant la "Pyramide Réduite Différentielle". Une application pour la description de visages est développée. Des résultats de classification à partir du coefficient de forme et utilisant une version adaptée des Machines à Vecteurs de Support, sur des bases de données issues des applications du monde réel sont présentés et comparés avec d'autres méthodes de classement basées sur des dissimilitudes. Il en ressort une forte robustesse de la méthode proposée avec des perfommances supérieures ou égales aux algorithmes de l'état de l'art. / The dissimilarity representation is an alternative for the use of features in the recognition of real world objects like images, spectra and time-signal. Instead of an absolute characterization of objects by a set of features, the expert or the system is asked to define a measure that estimates the dissimilarity between pairs of objects. Such a measure may also be defined for structural representations such as strings and graphs. The dissimilarity representation is potentially able to bridge structural and statistical pattern recognition. In this thesis we introduce a new fast Mahalanobis-like metric the “Shape Coefficient” for classification of dissimilarity data. Our approach is inspired by the Geometrical Discriminant Analysis and we have defined decision rules to mimic the behavior of the linear and quadratic classifier. The number of parameters is limited (two per class). We also expand and ameliorate this advantageous and rapid adaptive approach to learn only from dissimilarity representations by using the effectiveness of the Support Vector Machines classifier for real-world classification tasks. Several methods for incorporating dissimilarity representations are presented, investigated and compared to the “Shape Coefficient” in this thesis: • Pekalska and Duin prototype dissimilarity based classifiers; • Haasdonk's kernel based SVM classifier; • KNN classifier. Numerical experiments on artificial and real data show interesting behavior compared to Support Vector Machines and to KNN classifier: (a) lower or equivalent error rate, (b) equivalent CPU time, (c) more robustness with sparse dissimilarity data. The experimental results on real world dissimilarity databases show that the “Shape Coefficient” can be an alternative approach to these known methods and can be as effective as them in terms of accuracy for classification.

This paper introduced a Computer-Aided Optically Scanned Document Information Extraction System. It could extract information including invoice No., issued date, buyer, etc., from the optically scanned document to meet the demand of customs declaration companies. The system output the structured information to a relational database. In detail, a software architecture for the information extraction of diverse-structure optically scanned document is designed. In this system, the original document is classified firstly. It would put into template-based extraction to improve the extraction performance if its template is pre-defined in the system. Then, a method for image enhancement to improve the image classification is proposed. This method aims to optimize the accuracy of neural network model by extracting the template-related feature and actively removing the unrelated feature. Lastly, the above system is implemented in this paper. This extraction are programed in Python which is a cross-platform languages. This system comprises three parts, classification module, template-based extraction and non-template extraction all of which have APIs and could be ran independently. This feature make this system flexible and easy to customization for the further demand. 445 real-world customs document images were input to evaluate the system. The result revealed that the introduced system ensured the diverse document support with non-template extraction and reached the overall high performance with template-based extraction showing the goal was basically achieved.

In the forestry industry tree trunks are currently classified manually. The object of this thesis is to answer whether it is possible to automate this using modern computer hardware and image-classification of tree-trunks using machine learning algorithms. The report concludes, based on results from controlled experiments that it is possible to achieve an accuracy above 90% across the genuses Birch, Pine and Spruce with a classification-time per tree shorter than 500 milli seconds. The report further compares these results against previous research and concludes that better results are probable.

No description available.

No description available.

L’analyse des images satellite et aériennes figure parmi les sujets fondamentaux du domaine de la télédétection. Ces dernières années, les avancées technologiques ont permis d’augmenter la disponibilité à large échelle des images, en comprenant parfois de larges étendues de terre à haute résolution spatiale. En plus des questions évidentes de complexité calculatoire qui en surgissent, un de plus importants défis est l’énorme variabilité des objets dans les différentes régions de la terre. Pour aborder cela, il est nécessaire de concevoir des méthodes de classification qui dépassent l’analyse du spectre individuel de chaque pixel, en introduisant de l’information contextuelle de haut niveau. Dans cette thèse, nous proposons d’abord une méthode pour la classification avec des contraintes de forme, basée sur l’optimisation d’une structure de subdivision hiérarchique des images. Nous explorons ensuite l’utilisation des réseaux de neurones convolutionnels (CNN), qui nous permettent d’apprendre des descripteurs hiérarchiques profonds. Nous étudions les CNN depuis de nombreux points de vue, ce qui nous permettra de les adapter à notre objectif. Parmi les sujets abordés, nous proposons différentes solutions pour générer des cartes de classification à haute résolution et nous étudions aussi la récolte des données d’entrainement. Nous avons également créé une base de données d’images aériennes sur des zones variées, pour évaluer la capacité de généralisation des CNN. Finalement, nous proposons une méthode pour polygonaliser les cartes de classification issues des réseaux de neurones, afin de pouvoir les intégrer dans des systèmes d’information géographique. Au long de la thèse, nous conduisons des expériences sur des images hyperspectrales, satellites et aériennes, toujours avec l’intention de proposer des méthodes applicables, généralisables et qui passent à l’échelle. / The analysis of airborne and satellite images is one of the core subjects in remote sensing. In recent years, technological developments have facilitated the availability of large-scale sources of data, which cover significant extents of the earth’s surface, often at impressive spatial resolutions. In addition to the evident computational complexity issues that arise, one of the current challenges is to handle the variability in the appearance of the objects across different geographic regions. For this, it is necessary to design classification methods that go beyond the analysis of individual pixel spectra, introducing higher-level contextual information in the process. In this thesis, we first propose a method to perform classification with shape priors, based on the optimization of a hierarchical subdivision data structure. We then delve into the use of the increasingly popular convolutional neural networks (CNNs) to learn deep hierarchical contextual features. We investigate CNNs from multiple angles, in order to address the different points required to adapt them to our problem. Among other subjects, we propose different solutions to output high-resolution classification maps and we study the acquisition of training data. We also created a dataset of aerial images over dissimilar locations, and assess the generalization capabilities of CNNs. Finally, we propose a technique to polygonize the output classification maps, so as to integrate them into operational geographic information systems, thus completing the typical processing pipeline observed in a wide number of applications. Throughout this thesis, we experiment on hyperspectral, atellite and aerial images, with scalability, generalization and applicability goals in mind.

Machine learning is a growing area of artificial intelligence that is widely used in our world today. Training machine learning models requires knowledge and computing power. Machine Learning as a Service (MLaaS) tries to solve these issues. By storing the datasets and using virtual computing instances in the cloud, one can create machine learning models without writing a single line of code. When selecting an MLaaS platform to use, the natural question of which one to use arises. This thesis conducts controlled experiments to compare the image classification capabilities of Microsoft Azure ML, Amazon Web Services SageMaker, and Google Cloud Platform Vertex AI. The prediction accuracy, training time, and cost will be measured with three different datasets. Some subjective comments about the user experience while conducting these experiments will also be provided. The results of these experiments will be used to make recommendations as to which MLaaS platform to use depending on which metric is most suitable. This thesis found that Microsoft Azure ML performed best in terms of prediction accuracy, and training cost, across all datasets. Amazon Web Services SageMaker had the shortest time to train but performed the worst in terms of accuracy and had trouble with two of the three datasets. Google Cloud Platform Vertex AI did achieve the second-bestprediction accuracy but was the most expensive platform by far as it had the largest time to train. It did, however, provide the smoothest user experience.Overall, Azure ML would be the platform of choice for image classification tasks after weighing together the results of the experiments as well as the subjective user experience.

IoT-enheter blir allt populärare i takt med att de blir kraftfullare och skalbara. Därför var det viktigt att undersöka prestandan hos IoT-enheter när det kommer till maskininlärning. Huvudsyftet med detta arbete är att implementera två maskininlärningsmodeller på en bärbar dator och en Raspberry Pi 4 för att bestämma vilka maskininlärningsuppgifter som kan utföras på dessa enheter genom att sätta upp scenarier där vi kan testa låg-, medel- och högkrävande maskininlärning uppgifter på båda enheterna, som också möjliggör noggrann mätning. En bärbar dator användes som referenspunkt för att se om Raspberry Pi 4 kan prestera bra jämfört med en modern bärbar dator när de utför maskininlärningsuppgifter. Tensorflow valdes att användas för att implementera de två maskininlärningsmodellerna och för att mäta processen. Noggrannheten och träningstiden mättes för att bestämma prestandan för modellerna på de två enheterna. Tre datauppsättningar valdes ut för att användas för att träna och testa modellerna på de två enheterna, dessa datauppsättningar innehöll bilder, den första datauppsättningen bestod av mycket små bilder, och den andra bestod av lite större bilder, den sista datauppsättningen bestod av ännu större bilder, detta gjordes för att testa tre olika svårighetsgrader för modellerna på de två olika enheterna. Efter att träningen och utvärdering av modellerna slutförts på båda enheterna med hjälp av de tre datauppsättningarna analyserades de resulterande mätningarna och diskuterades sedan för att nå en slutsats. Slutsatsen från detta arbete var att endast lågnivåmaskininlärningsuppgifter är ett gångbart alternativ nu på grund av den extrema tid som krävs för att träna modellen för bildklassificering, men om tiden inte är en viktig faktor, skulle Raspberry Pi 4 efter en lång tid fortfarande uppnå samma noggrannhet som den bärbara datorn gjorde. / IoT devices are becoming increasingly popular as they are becoming more powerful and scalable. Therefor it was important to examine the potential of IoT devices when it comes to Machine Learning. The main objective of this work is to implement two machine learning models on a laptop and a Raspberry Pi 4 to determine what machine learning tasks that can be performed on these devices by setting up scenarios where we can test low, mid and high demanding machine learning tasks on both devices that also allows for accurate measurement being taken. A laptop was used a reference point to see if the Raspberry Pi 4 can perform well relative to a modern-day laptop when performing machine learning tasks. Tensorflow was chosen to be used to implement the two machine learning models and to measure the process. The accuracy and training time were measured to determine the performance of the models on the two devices. Three datasets were chosen to be used for training and testing the models on the two devices, these datasets contained images, the first dataset consisted of very tiny images, and the second one consisted of slightly larger images, the last dataset consisted of very large images, this was done to test three different levels of difficulty for the models. After training and evaluation of the models were completed on both devices using the three datasets, the resulting measurements were analyzed and then discussed to reach the conclusion. The conclusion from this work was that only low-tier machine learning tasks are a viable option now because of the extreme amount of time required to train the model for image classification, however if time isn’t an important factor, the Raspberry Pi 4 would after a long time still reach the same accuracy as the laptop did.

Image classification is a memory- and compute-intensive task. It is difficult to implement high-speed image classification algorithms on resource-limited systems like FPGAs and embedded computers. Most image classification algorithms require many fixed- and/or floating-point operations and values. In this work, we explore the use of binary values to reduce the memory and compute requirements of image classification algorithms. Our objective was to implement these algorithms on resource-limited systems while maintaining comparable accuracy and high speeds. By implementing high-speed image classification algorithms on resource-limited systems like embedded computers, FPGAs, and ASICs, automated visual inspection can be performed on small low-powered systems. Industries like manufacturing, medicine, and agriculture can benefit from compact, high-speed, low-power visual inspection systems. Tasks like defect detection in manufactured products and quality sorting of harvested produce can be performed cheaper and more quickly. In this work, we present ECO Jet Features, an algorithm adapted to use binary values for visual inspection. The ECO Jet Features algorithm ran 3.7x faster than the original ECO Features algorithm on embedded computers. It also allowed the algorithm to be implemented on an FPGA, achieving 78x speedup over full-sized desktop systems, using a fraction of the power and space. We reviewed Binarized Neural Nets (BNNs), neural networks that use binary values for weights and activations. These networks are particularly well suited for FPGA implementation and we compared and contrasted various FPGA implementations found throughout the literature. Finally, we combined the deep learning methods used in BNNs with the efficiency of Jet Features to make Neural Jet Features. Neural Jet Features are binarized convolutional layers that are learned through deep learning and learn classic computer vision kernels like the Gaussian and Sobel kernels. These kernels are efficiently computed as a group and their outputs can be reused when forming output channels. They performed just as well as BNN convolutions on visual inspection tasks and are more stable when trained on small models.

Indiana University-Purdue University Indianapolis (IUPUI) / Every year the most effective Deep learning models, CNN architectures are showcased based on their compatibility and performance on the embedded edge hardware, especially for applications like image classification. These deep learning models necessitate a significant amount of computation and memory, so they can only be used on high-performance computing systems like CPUs or GPUs. However, they often struggle to fulfill portable specifications due to resource, energy, and real-time constraints. Hardware accelerators have recently been designed to provide the computational resources that AI and machine learning tools need. These edge accelerators have high-performance hardware which helps maintain the precision needed to accomplish this mission. Furthermore, this classification dilemma that investigates channel interdependencies using either depth-wise or group-wise convolutional features, has benefited from the inclusion of Bottleneck modules. Because of its increasing use in portable applications, the classic inverted residual block, a well-known architecture technique, has gotten more recognition. This work takes it a step forward by introducing a design method for porting CNNs to lowresource embedded systems, essentially bridging the difference between deep learning models and embedded edge systems. To achieve these goals, we use closer computing strategies to reduce the computer’s computational load and memory usage while retaining excellent deployment efficiency. This thesis work introduces HBONext, a mutated version of Harmonious Bottlenecks (DHbneck) combined with a Flipped version of Inverted Residual (FIR), which outperforms the current HBONet architecture in terms of accuracy and model size miniaturization. Unlike the current definition of inverted residual, this FIR block performs identity mapping and spatial transformation at its higher dimensions. The HBO solution, on the other hand, focuses on two orthogonal dimensions: spatial (H/W) contraction-expansion and later channel (C) expansion-contraction, which are both organized in a bilaterally symmetric manner. HBONext is one of those versions that was designed specifically for embedded and mobile applications. In this research work, we also show how to use NXP Bluebox 2.0 to build a real-time HBONext image classifier. The integration of the model into this hardware has been a big hit owing to the limited model size of 3 MB. The model was trained and validated using CIFAR10 dataset, which performed exceptionally well due to its smaller size and higher accuracy. The validation accuracy of the baseline HBONet architecture is 80.97%, and the model is 22 MB in size. The proposed architecture HBONext variants, on the other hand, gave a higher validation accuracy of 89.70% and a model size of 3.00 MB measured using the number of parameters. The performance metrics of HBONext architecture and its various variants are compared in the following chapters.