Framgången inom maskininlärning och djupinlärning beror till stor del på stora, annoterade dataset. Att tilldela etiketter till data är väldigt resurskrävande och kan till viss del undvikas genom att utnyttja datans statistiska egenskaper. En maskininlärningsmodell kan lära sig att klassificera bilder från en domän utifrån träningsexempel som innehåller bilder, samt etiketter som berättar vad bilder föreställer. Men vad gör man om datan inte har tilldelade etiketter? En maskininlärningsmodell som lär sig en uppgift utifrån annoterad data från en källdomän, kan med hjälp av information från måldomänen (som inte har tilldelade etiketter), anpassas till att prestera bättre på data från måldomänen. Forskningsområdet som studerar hur man anpassar och generaliserar en modell mellan två olika domäner heter domänanpassning, eller domain adaptation, på engelska.   Detta examensarbete är utfört på Scanias forskningsavdelning för autonom transport och handlar om hur modeller för bildklassificering som tränas på kamerabilder med etiketter, kan anpassas till att få ökad noggrannhet på ett dataset med LiDAR bilder, som inte har etiketter. Två metoder för domänanpassning har jämförts med varandra, samt en model tränad på kameradata genom övervakad inlärning utan domänanpassning. Alla metoder opererar på något vis med ett djupt faltningsnätverk (CNN) där uppgiften är att klassificera bilder utav bilar eller fotgängare. Kovariansen utav datan från käll- och måldomänen är det centrala måttet för domänanpassningsmetoderna i detta projekt. Den första metoden är en så kallad ytlig metod, där själva anpassningsmetoden inte ingår inuti den djupa arkitekturen av modellen, utan är ett mellansteg i processen. Den andra metoden förenar domänanpassningsmetoden med klassificeringen i den djupa arkitekturen. Den tredje modellen består endast utav faltningsnätverket, utan en metod för domänanpassning och används som referens.    Modellen som tränades på kamerabilderna utan en domänanpassningsmetod klassificerar LiDAR-bilderna med en noggrannhet på 63.80%, samtidigt som den ”ytliga” metoden når en noggrannhet på 74.67% och den djupa metoden presterar bäst med 80.73%. Resultaten visar att det är möjligt att anpassa en modell som tränas på data från källdomänen, till att få ökad klassificeringsnoggrannhet i måldomänen genom att använda kovariansen utav datan från de två domänerna. Den djupa metoden för domänanpassning tillåter även användandet utav andra statistiska mått som kan vara mer framgångsrika i att generalisera modellen, beroende på hur datan är fördelad. Överlägsenheten hos den djupa metoden antyder att domänanpassning med fördel kan bäddas in i den djupa arkitekturen så att modelparametrarna blir uppdaterade för att lära sig en mer robust representation utav måldomänen.

[pt] Modelos baseados em redes neurais profundas como as Redes Neurais Convolucionais proporcionaram avanços significativos em diversas áreas da computação. No entanto, essa tecnologia é ainda pouco aplicada à predição de qualidade sísmica, que é uma atividade relevante para exploração de hidrocarbonetos. Ser capaz de, rapidamente, classificar o ruído presente em aquisições de dados sísmicos permite aceitar ou rejeitar essas aquisições de forma eficiente, o que além de economizar recursos também melhora a interpretabilidade dos dados. Neste trabalho apresenta-se um dataset criado a partir de 6.918 aquisições manualmente classificadas pela percepção de especialistas e pesquisadores, que serviu de base para o treinamento, validação e testes de um classificador, também proposto neste trabalho, baseado em uma rede neural convolucional. Em resultados empíricos, observou-se-se um F1 Score de 95,58 porcento em uma validação cruzada de 10 folds e 93,56 porcento em um conjunto de holdout de teste. / [en] Deep Learning based models, such as Convolutional Neural Networks (CNNs), have led to significant advances in several areas of computing applications. Nevertheless, this technology is still rarely applied to seismic quality prediction, which is a relevant task in hydrocarbon exploration. Being able to promptly classify noise in common shot gather(CSG) acquisitions of seismic data allows the acceptance or rejection of those aquisitions, not only saving resources but also increasing the interpretability of data. In this work, we introduce a real-world classification dataset based on 6.918 common shot gather, manually labeled by perception of specialists and researches. We use it to train a CNN classification model for seismic shot-gathers quality prediction. In our empirical evaluation, we observed an F1 Score of 95,58 percent in 10 fold cross-validation and 93,56 percent in a Holdout Test.

<p dir="ltr">In the dynamic and ever-evolving landscape of Artificial Intelligence (AI), the domain of deep learning has emerged as a pivotal force, propelling advancements across a broad spectrum of applications, notably in the intricate field of image classification. Image classification, a critical task that involves categorizing images into predefined classes, serves as the backbone for numerous cutting-edge technologies, including but not limited to, automated surveillance, facial recognition systems, and advanced diagnostics in healthcare. Despite the significant strides made in the area, the quest for models that not only excel in accuracy but also demonstrate robust generalization across varied datasets, and maintain resilience against the pitfalls of overfitting, remains a formidable challenge.</p><p dir="ltr">EfficientNetB0, a model celebrated for its optimized balance between computational efficiency and accuracy, stands at the forefront of solutions addressing these challenges. However, the nuanced complexities of datasets such as CIFAR-10, characterized by its diverse array of images spanning ten distinct categories, call for specialized adaptations to harness the full potential of such sophisticated architectures. In response, this thesis introduces an optimized version of the EffciientNetB0 architecture, meticulously enhanced with strategic architectural modifications, including the incorporation of an additional Dense layer endowed with 512 units and the strategic use of Dropout regularization. These adjustments are designed to amplify the model's capacity for learning and interpreting complex patterns inherent in the data.</p><p dir="ltr">Complimenting these architectural refinements, a nuanced two-phase training methodology is also adopted in the proposed model. This approach commences with the initial phase of training where the base model's pre-trained weights are frozen, thus leveraging the power of transfer learning to secure a solid foundational understanding. The subsequent phase of fine-tuning, characterized by the selective unfreezing of layers, meticulously calibrates the model to the intricacies of the CIFAR-10 dataset. This is further bolstered by the implementation of adaptive learning rate adjustments, ensuring the model’s training process is both efficient and responsive to the nuances of the learning curve.</p><p><br></p>

No / The bag of visual words (BOW) model is an efficient image representation technique for image categorization and annotation tasks. Building good visual vocabularies, from automatically extracted image feature vectors, produces discriminative visual words, which can improve the accuracy of image categorization tasks. Most approaches that use the BOW model in categorizing images ignore useful information that can be obtained from image classes to build visual vocabularies. Moreover, most BOW models use intensity features extracted from local regions and disregard colour information, which is an important characteristic of any natural scene image. In this paper, we show that integrating visual vocabularies generated from each image category improves the BOW image representation and improves accuracy in natural scene image classification. We use a keypoint density-based weighting method to combine the BOW representation with image colour information on a spatial pyramid layout. In addition, we show that visual vocabularies generated from training images of one scene image dataset can plausibly represent another scene image dataset on the same domain. This helps in reducing time and effort needed to build new visual vocabularies. The proposed approach is evaluated over three well-known scene classification datasets with 6, 8 and 15 scene categories, respectively, using 10-fold cross-validation. The experimental results, using support vector machines with histogram intersection kernel, show that the proposed approach outperforms baseline methods such as Gist features, rgbSIFT features and different configurations of the BOW model.

Modern deep learning models require large amounts of data to train, and the acquisition of data can be challenging. Synthetic data provides an alternative to manually collecting real data, alleviating problems associated with real data acquisition. For recycling processes, classifying metal scrap piles containing hazardous objects is important, where hazardous objects can be damaging and costly if handled incorrectly. Automatically detecting hazardous objects in metal scrap piles using image classification models requires large amounts of data, and metal scrap piles contain large variations in objects, textures, and lighting. Furthermore, data acquisition can be challenging in the recycling domain, where positive objects can be scarce and manual acquisition setup can be challenging. In this thesis, synthetic images of metal scrap piles in a recycling process are created, intended for training image classification models to detect metal scrap piles containing fire extinguishers or hydraulic cylinders. Synthetic images are created with physically based rendering and domain randomization, rendered with either rasterization or ray tracing engines. Ablation studies are conducted to investigate the effect of using domain randomization. The performance of models trained on purely synthetic datasets is compared to models trained on datasets containing only real images. Furthermore, photorealistic rendering with ray tracing rendering is evaluated by comparing F1 scores between models trained on data sets created with rasterization or ray tracing. The F1 scores show that models trained on purely synthetic data outperform those trained solely on real data when classifying images containing fire extinguishers or hydraulic cylinders. Ablation studies show that domain randomization of textures is beneficial both for the classification of fire extinguishers and for the classification of hydraulic cylinders in metal scrap piles. High dynamic range image lighting randomization does not provide benefits when classifying metal scrap piles containing fire extinguishers, suggesting that other lighting randomization techniques may be more effective. The F1 scores show that synthetically created images using rasterization perform better when classifying metal scrap piles containing fire extinguishers. However, when classifying metal scrap piles containing hydraulic cylinders, images created with ray tracing achieve higher F1 scores. This thesis highlights the potential of synthetic data as an alternative to manually acquiring real data, particularly in domains where data collection is challenging and time-consuming. The results show the effectiveness of domain randomization and physically based rendering techniques in creating realistic and diverse synthetic datasets.

完善的崩塌地資料庫有助於地區土地利用的適宜性評估、與環境保護措施之研訂。目前，崩塌地地理空間資料(Geospatial data)的產生方法主要為：人為判釋經正射糾正(Ortho-rectification)的遙測影像，基於該影像，將辨識目標數位化(Digitizing)。然而，遙測影像的「正射糾正」與「人為判釋」往往不敷災後的緊急需求。為促進資料收集效率，本研究試圖發展一套自動化流程：以「未經糾正的遙測影像」為判釋對象，判釋作業以「物件式影像分類(Object-based classification)技術」進行，並利用「現存地形資料」，實現自動判釋所產生之辨識成果的地理對位(Georeferencing)與過濾篩選；最後，以「與現存各類輔助資料的套疊分析成果」為其屬性，以便利崩塌地地理空間資料的後續應用。 物件式影像分類分為為「影像分割(Image segmentation)」與「物件分類」兩步驟。於影像分割階段，採用多重解析度分割法(Multiresolution segmentation algorithm)─由於陰影下各類地物的影像光譜差異較不明顯，為避免陰影區之錯誤分割，賦予陰影區較小的尺度參數(Scale parameter)；於物件分類階段，基於訓練資料，以「線性核函數的支持向量機(Support Vector Machine, SVM, with a linear kernel)」為分類器，偵測「非雲與植被區」，並輸出為向量式資料(Vector data)。而後基於現存地形資料，以光線追蹤法(Ray-tracing algorithm)進行分類器輸出向量式資料的地理對位，並自訂第二階段的地形特徵過濾準則。實驗成果顯示，此自動化流程產出的崩塌地地理空間資料─其生產者精度(Producer’s accuracy)與使用者精度(User’s accuracy)分別介於0.85~0.99與0.44~0.96。

<p>The Convolutional Neural Network (CNN) have shown a substantial improvement in the field of Machine Learning. But they do come with their own set of drawbacks. Capsule Networks have addressed the limitations of CNNs and have shown a great improvement by calculating the pose and transformation of the image. Deeper networks are more powerful than shallow networks but at the same time, more difficult to train. Residual Networks ease the training and have shown evidence that they can give good accuracy with considerable depth. Putting the best of Capsule Network and Residual Network together, we present Residual Capsule Network and 3-Level Residual Capsule Network, a framework that uses the best of Residual Networks and Capsule Networks. The conventional Convolutional layer in Capsule Network is replaced by skip connections like the Residual Networks to decrease the complexity of the Baseline Capsule Network and seven ensemble Capsule Network. We trained our models on MNIST and CIFAR-10 datasets and have seen a significant decrease in the number of parameters when compared to the Baseline models.</p>

La représentation d'image est au cœur de beaucoup d'algorithmes de vision par ordinateur. Elle intervient notamment dans des tâches de reconnaissance de catégories visuelles comme la classification ou la détection d'objets. Dans ce contexte, la représentation "sac de mot visuel" (Bag of Visual Words ou BoVW en anglais) est l'une des méthodes de référence. Dans cette thèse, nous nous appuyons sur ce modèle pour proposer des représentations d'images discriminantes. Dans la première partie, nous présentons une nouvelle approche simple et efficace pour prendre en compte des informations spatiales dans le modèle BoVW. Son principe est de considérer l'orientation et la longueur de segments formés par des paires de descripteurs similaires. Une notion de "softsimilarité" est introduite pour définir ces relations intra et inter mots visuels. Nous montrons expérimentalement que notre méthode ajoute une information discriminante importante au modèle BoVW et que cette information est complémentaire aux méthodes de l'état de l'art. Ensuite, nous nous focalisons sur la description de l'information couleur. Contrairement aux approches traditionnelles qui s'appuient sur des descriptions invariantes aux changements d'éclairage, nous proposons un descripteur basé sur le pouvoir discriminant. Nos expérimentations permettent de conclure que ce descripteur apprend automatiquement un certain degré d'invariance photométrique tout en surclassant les descripteurs basés sur cette invariance photométrique. De plus, combiné avec un descripteur de forme, le descripteur proposé donne des résultats excellents sur quatre jeux de données particulièrement difficiles. Enfin, nous nous intéressons à la représentation de la couleur à partir de la réflectance multispectrale des surfaces observées, information difficile à extraire sans instruments sophistiqués. Ainsi, nous proposons d'utiliser l'écran et la caméra d'un appareil portable pour capturer des images éclairées par les couleurs primaires de l'écran. Trois éclairages et trois réponses de caméra produisent neuf valeurs pour estimer la réflectance. Les résultats montrent que la précision de la reconstruction spectrale est meilleure que celle estimée avec un seul éclairage. Nous concluons que ce type d'acquisition est possible avec des appareils grand public tels que les tablettes, téléphones ou ordinateurs portables / Image representation is in the heart of many computer vision algorithms. Different computer vision tasks (e.g. classification, detection) require discriminative image representations to recognize visual categories. In a nutshell, the bag-of-visual-words image representation is the most successful approach for object and scene recognition. In this thesis, we mainly revolve around this model and search for discriminative image representations. In the first part, we present a novel approach to incorporate spatial information in the BoVW method. In this framework, we present a simple and efficient way to infuse spatial information by taking advantage of the orientation and length of the segments formed by pairs of similar descriptors. We introduce the notion of soft-similarity to compute intra and inter visual word spatial relationships. We show experimentally that, our method adds important discriminative information to the BoVW method and complementary to the state-of-the-art method. Next, we focus on color description in general. Differing from traditional approaches of invariant description to account for photometric changes, we propose discriminative color descriptor. We demonstrate that such a color description automatically learns a certain degree of photometric invariance. Experiments show that the proposed descriptor outperforms existing photometric invariants. Furthermore, we show that combined with shape descriptor, the proposed color descriptor obtain excellent results on four challenging data sets.Finally, we focus on the most accurate color representation i.e. multispectral reflectance which is an intrinsic property of a surface. Even with the modern era technological advancement, it is difficult to extract reflectance information without sophisticated instruments. To this end, we propose to use the display of the device as an illuminant while the camera captures images illuminated by the red, green and blue primaries of the display. Three illuminants and three response functions of the camera lead to nine response values which are used for reflectance estimation. Results show that the accuracy of the spectral reconstruction improves significantly over the spectral reconstruction based on a single illuminant. We conclude that, multispectral data acquisition is potentially possible with consumer hand-held devices such as tablets, mobiles, and laptops

Em países em desenvolvimento, doenças infecciosas se constituem ainda um grave problema de saúde pública. Muitas vezes, essas doenças estão altamente relacionadas a condições urbanas que podem ser encontradas em áreas mais pobres. Nesses casos, o sensoriamento remoto (SR) pode ser utilizado como uma poderosa ferramenta de estudo. Novos produtos de SR se encontram disponíveis no mercado, permitindo o desenvolvimento de análises espaciais cada vez mais profundas e precisas. No entanto, a complexidade que envolve a epidemiologia de doenças, a irregularidade de ocupações urbanas e a heterogeneidade das imagens de alta resolução espacial têm restringido o desenvolvimento de estudos nesse campo científico. O desafio de identificar elementos e tipologias urbanas em imagens de sensoriamento remoto relacionadas à ocorrência da leptospirose justifica-se pela crença de que ferramentas de SR podem ser mais amplamente utilizadas no monitoramento de carências urbanísticas e, consequentemente, na gestão de ações e investimentos públicos. A metodologia contempla uma revisão bibliográfica sistemática, com base na qual foram criados modelos de transmissão da leptospirose e investigadas tipologias urbanas presentes na área de estudo. As variáveis baseadas em dados de SR que formam os indicadores dos modelos e que caracterizam as tipologias foram usadas para definir objetos e atributos, alvos das investigações em imagens de alta resolução espacial. Os procedimentos de SR adotados baseiam-se na segmentação multi-nível, classificação baseada em objeto, e utilizam ortofotografias aéreas, imagem QuickBird e base cartográfica do eixo viário do Subúrbio Ferroviário de Salvador. Para o cálculo das variáveis utilizou-se produtos do processamento da imagem QuickBird. Procedimentos de geoprocessamento foram realizados em sistema de informações geográficas. Por fim, realizaram-se as primeiras análises epidemiológicas que investigam a relação da leptospirose com os elementos e tipologias urbanas identificados por meio de SR, cujos resultados apontam maior influência do percentual de pavimentação das vias, sua largura e qualidade da edificação na possibilidade de ocorrência da leptospirose no Subúrbio. Possíveis fontes de viés são discutidas ao lado de propostas de continuação da pesquisa. Apesar dos problemas e limitações identificados no processo, o estudo mostra que a metodologia desenvolvida baseada em SR se constitui uma poderosa ferramenta de análise do espaço intra-urbano, uma vez que permite a identificação de elementos e tipologias relacionados a situações de risco, apoiando assim, o direcionamento de investimentos públicos que venham refletir na melhoria das condições de saúde da população. / In developing countries, infectious diseases are still a serious public health problem. These diseases are often and highly related to urban conditions found in poor areas, in these cases, remote sensing (RS) can be used as a powerful tool. New RS products are now available allowing the development of more complex and precise spatial analysis. On the other hand, the complexity of epidemiological studies, the lack of regularity of precarious urban settlements and the heterogeneity of high spatial resolution images have been restricting the development of studies in this areas. The challenge of identifying urban elements and typologies related to the leptospirosis using RS products is pursued due the belief that RS can be more used among professionals and researchers in the task of monitoring the urban environment, and directing public investments and actions. The methodology presented consists in a broad literature review, which was used to support leptospirosis transmission risk models and to find urban typologies at the study area. Variables based on RS were identified in the disease models and in the typologies characterization. This models and typologies also defined targets to look for in the high spatial resolution images. RS procedures were based on multi-level segmentation, object-based classification, aerial photography, QuickBird satellite images and street axis vector data of the Railroad Suburb of Salvador. In order to obtain the variable\'s values, results of QuickBird image processing were added to a geographic database and processed using vector and raster over layering techniques. At last, epidemiological analysis were initiated aiming to find its relationship with the urban elements and typologies identified using RS. The results points paved streets, streets wideness and house quality as the RS variables that have more influence on the leptospirosis transmission chance. The dissertation also presents research restrains, potentials, possible sources of bias and future studies proposals. It concludes that the RS based methodology presented is a powerful tool for urban analysis, due to its capabilities for identifying urban targets related to risky situations, and, therefore, for helping direct public investments to improve life conditions an unprivileged city areas.

O objetivo desta Tese é a proposição de uma metodologia alternativa para planejamento de transportes que contempla as ferramentas disponíveis na ciência do sensoriamento remoto. A perspectiva adotada analisa aspectos do planejamento de transportes urbanos, tendo como embasamento os dados e informações advindos das imagens de satélite com alto poder de resolução espacial. A metodologia usa a abordagem baseada em objetos para classificar imagens de satélite de sensoriamento remoto. Através do processo de classificação, identificam-se feições urbanas úteis para o planejamento de transporte, em especial áreas de concentração de polos geradores de viagens do município de João Pessoa no estado da Paraíba, Brasil. A proposta é que com base nesses dados, e outros provenientes de uma pesquisa de campo (pesquisa domiciliar origem/destino), é possível caracterizar o uso do solo e a correspondente demanda por transportes. O estudo se justifica por propor uma alternativa mais ágil e menos onerosa, em comparação aos métodos tradicionais de construção e atualização da base de dados para análises de transportes. Ao identificar as regiões da cidade com as maiores quantidades de viagens geradas, os resultados obtidos auxiliam nas ações de planejamento do sistema de transportes, visando alcançar o equilíbrio entre oferta e demanda de transporte com o uso do solo urbano. / The objective of this Thesis is to propose an alternative method of transportation planning that considers the tools available in the science of remote sensing. The perspective adopted examines aspects of urban transportation planning, having as basis the data and information coming from satellite images with high spatial resolution. The methodology uses the object-based approach to classify remote sensing satellite imagery. Through the classification process, urban features useful for transportation planning are identified, mainly areas of concentration of trip generation in the city of João Pessoa, state of Paraíba, Brazil. The proposal is that, based on these data, and others from a field research (origin/destination home-interview survey), it is possible to characterize the land use and the corresponding demand for transport. The study is justified because it proposes a more agile and less costly alternative, compared to traditional methods of building and updating the database for transport analysis. By identifying areas of the city with the largest amounts of trips generated, the results support planning actions on the transportation system, in order to achieve a balance between transport supply and demand with urban land use.