Under de senaste året har modeller för djupinlärning använts inom nästa alla områden, från industri till akademi, särskilt för bildklassifikation. Dessa modeller är dock enorma i storlek, med miljontals parametrar, vilket gör det svårt att distribuera till mindre enheter med begränsade resurser såsom mobiltelefoner. Denna studie tar upp små modeller av faltningsnätverk som är toppmoderna inom djupinlärning och vars storlek är lämplig för mobilapplikation. Syftet med denna studie är att utvärdera prestanda på faltningsnätverken Xception, InceptionV3 och MobilNetV2 för att underlätta vid valbeslut av faltningsnätverk som bas vid utveckling av mobila applikation inom bildklassificering. För att uppnå syftet har dessa faltningsnätverk implementeras med hjälp av överföringsinlärning metod samt utformas för att skilja på bilder av nätpaneler från företaget Troax. Studien tar upp metoden som möjliggör att överföra kunskap från befintliga förtränade modeller till nya modeller. Studien förklarar även hur träningsprocessen och testprocessen gick till samt analys kring resultatet.   Resultat visade att Xception hade 86 % noggrannhet med en processtid på 10 minuter på 2000 träningsbilder och 1000st testbilder. Xceptions prestation var bäst bland alla dessa modeller. Skillnaden mellan Xception och Inception var på 10 % noggrannhet och 2 minuter processtid. Mellan Xception och MobilNetV2 var skillnaden på 23 % noggrannhet och 3 minuter processtid. Experimentet visade att dessa modeller presterade mindre bra vid mindre träningsbilder under 800st. Över 800st bilder började respektive modell att utföra prediktering över 70 % noggrannhet. / In recent years, deep learning models have been used in almost all areas, from industry to academia, specifically for image classification. However, these models are huge in size, with millions of parameters, making it difficult to distribute to smaller devices with limited resources such as mobile phones. This study addresses lightweight pre-trained models of convolutional neural networks which is state of art in deep learning and their size is suitable as a base model for mobile application development. The purpose of this study is to evaluate the performance of Xception, InceptionV3 and MobilNetV2 in order to facilitate selection decisions of a lightweight convolutional networks as base for the development of mobile applications in image classification. In order to achieve their purpose, these models have been implemented using the Transfer Learning method and are designed to distinguish images on mesh panels from the company Troax. The study takes up the method that allows transfer of knowledge from an existing model to a new model, explain how the training process and the test process went, as well as analysis of results. Results showed that Xception had 86% accuracy and had 10 minutes processing time on 2000 training images and 1000 test images. Exception’s performance was the best among all these models. The difference between Xception and InceptionV3 was 10% accuracy and 2 minutes process time. Between Xception and MobilNetV2 there was a difference of 23% in accuracy and 3 minutes in process time. Experiments showed that these models performed less well with smaller training images below 800 images. Over 800 images, each model began to perform prediction over 70% accuracy.

Research work presented within this thesis propose a neoteric variant of deep convolutional neural network architecture, CondenseNeXt, designed specifically for ARM-based embedded computing platforms with constrained computational resources. CondenseNeXt is an improved version of CondenseNet, the baseline architecture whose roots can be traced back to ResNet. CondeseNeXt replaces group convolutions in CondenseNet with depthwise separable convolutions and introduces group-wise pruning, a model compression technique, to prune (remove) redundant and insignificant elements that either are irrelevant or do not affect performance of the network upon disposition. Cardinality, a new dimension to the existing spatial dimensions, and class-balanced focal loss function, a weighting factor inversely proportional to the number of samples, has been incorporated in order to relieve the harsh effects of pruning, into the design of CondenseNeXt’s algorithm. Furthermore, extensive analyses of this novel CNN architecture was performed on three benchmarking image datasets: CIFAR-10, CIFAR-100 and ImageNet by deploying the trained weight on to an ARM-based embedded computing platform: NXP BlueBox 2.0, for real-time image classification. The outputs are observed in real-time in RTMaps Remote Studio’s console to verify the correctness of classes being predicted. CondenseNeXt achieves state-of-the-art image classification performance on three benchmark datasets including CIFAR-10 (4.79% top-1 error), CIFAR-100 (21.98% top-1 error) and ImageNet (7.91% single model, single crop top-5 error), and up to 59.98% reduction in forward FLOPs compared to CondenseNet. CondenseNeXt can also achieve a final trained model size of 2.9 MB, however at the cost of 2.26% in accuracy loss. Thus, performing image classification on ARM-Based computing platforms without requiring a CUDA enabled GPU support, with outstanding efficiency.<br>

Cílem této diplomové práce je automatizace detekce vad ve vláknitých materiálech. Firma SILON se již přes padesát let zabývá výrobou jemné vaty z recyklovaných PET lahví. Tato vata se následně používá ve stavebnictví, automobilovém průmyslu, ale nejčastěji v dámských hygienických potřebách a dětských plenách. Cílem firmy je produkovat co nejkvalitnější výrobek a proto je každá dávka testována v laboratoři s několika přísnými kritérii. Jednám z testů je i množství vadných vláken, jako jsou zacuchané smotky vláken, nebo nevydloužená vlákna, která jsou tvrdá a snadno se lámou. Navrhovaný systém sestává ze snímací lavice fungující jako scanner, která nasnímá vzorek vláken, který byl vložen mezi dvě skleněné desky. Byla provedena série testů s různým osvětlením, která ověřovala vlastnosti Rhodaminu, který se používá právě na rozlišení defektů od ostatních vláken. Tyto defekty mají zpravidla jinou molekulární strukturu, na kterou se barvivo chytá lépe. Protože je Rhodamin fluorescenční barvivo, je možné ho například pod UV světlem snáze rozeznat. Tento postup je využíván při manuální detekci. Při snímání kamerou je možno si vypomoci filtrem na kameře, který odfiltruje excitační světlo a propustí pouze světlo vyzářené Rhodaminem. Součástí výroby skeneru byla i tvorba ovládacího programu. Byla vytvořena vlastní knihovna pro ovládání motoru a byla upravena knihovna pro kameru. Oba systém pak bylo možno ovládat pomocí jednotného GUI, které zajišťovalo pořizování snímku celé desky. Pomocí skeneru byla nasnímána řada snímků, které bylo třeba anotovat, aby bylo možné naučit počítač rozlišovat defekty. Anotace proběhla na pixelové úrovni; každý defekt byl označen v grafickém editoru ve speciální vrstvě. Pro rozlišování byla použita umělá neuronová síť, která funguje na principu konvolucí. Tento typ sítě je navíc plně konvoluční, takže výstupem sítě je obraz, který by měl označit na tom původním vadné pixely. Výsledky naučené sítě jsou v práci prezentovány a diskutovány. Síť byla schopna se naučit rozeznávat většinu defektů a spolehlivě je umí rozeznat a segmentovat. Potíže má v současné době s detekcí rozmazaných defektů na krajích zorného pole a s defekty, jejichž hranice není tolik zřetelná na vstupních obrazech. Nutno zmínit, že zákazník má zájem o kompletní řešení scanneru i s detekčním softwarem a vývoj tohoto zařízení bude pokračovat i po závěru této diplomové práce.

The thesis focuses on modern methods of traffic sign detection and traffic lights detection directly in traffic and with use of back analysis. The main subject is convolutional neural networks (CNN). The solution is using convolutional neural networks of YOLO type. The main goal of this thesis is to achieve the greatest possible optimization of speed and accuracy of models. Examines suitable datasets. A number of datasets are used for training and testing. These are composed of real and synthetic data sets. For training and testing, the data were preprocessed using the Yolo mark tool. The training of the model was carried out at a computer center belonging to the virtual organization MetaCentrum VO. Due to the quantifiable evaluation of the detector quality, a program was created statistically and graphically showing its success with use of ROC curve and evaluation protocol COCO. In this thesis I created a model that achieved a success average rate of up to 81 %. The thesis shows the best choice of threshold across versions, sizes and IoU. Extension for mobile phones in TensorFlow Lite and Flutter have also been created.

Cílem této práce je vytvořit neuronovou síť která bude schopna rekonstruovat obličeje z fotografií na kterých je část obličeje překrytá maskou. Jsou prezentovány koncepty využívané při vývoji konvolučních neuronových sítí a generativních kompetitivních sítí. Dále jsou popsány koncepty používané v neuronových sítích specificky pro rekonstrukci fotografií obličejů. Je představen model generativní kompetitivní sítě využívající kombinaci hrazených konvolučních vrstev a víceškálových bloků schopný realisticky doplnit oblasti obličeje zakryté maskou.

Deep learning has shown to be successful on the task of semantic segmentation of three-dimensional (3D) point clouds, which has many interesting use cases in areas such as autonomous driving and defense applications. A common type of sensor used for collecting 3D point cloud data is Light Detection and Ranging (LiDAR) sensors. In this thesis, a time-correlated single-photon counting (TCSPC) LiDAR is used, which produces very accurate measurements over long distances up to several kilometers. The dataset collected by the TCSPC LiDAR used in the thesis contains two classes, person and other, and it comes with several challenges due to it being limited in terms of size and variation, as well as being extremely class imbalanced. The thesis aims to identify, analyze, and evaluate state-of-the-art deep learning models for semantic segmentation of point clouds produced by the TCSPC sensor. This is achieved by investigating different loss functions, data variations, and data augmentation techniques for a selected state-of-the-art deep learning architecture. The results showed that loss functions tailored for extremely imbalanced datasets performed the best with regard to the metric mean intersection over union (mIoU). Furthermore, an improvement in mIoU could be observed when some combinations of data augmentation techniques were employed. In general, the performance of the models varied heavily, with some achieving promising results and others achieving much worse results.

Implementing object detection methods for runway detection during landing approaches is limited in the safety-critical aircraft domain. This limitation is due to the difficulty that comes with verification of the design and the ability to understand how the object detection behaves during operation. During operation, object detection needs to consider the aircraft's position, environmental factors, different runways and aircraft attitudes. Training such an object detection model requires a comprehensive dataset that defines the features mentioned above. The feature's impact on the detection capabilities needs to be analysed to ensure the correct distribution of images in the dataset. Gathering images for these scenarios would be costly and needed due to the aviation industry's safety standards. Synthetic data can be used to limit the cost and time required to create a dataset where all features occur. By using synthesised data in the form of generating datasets in a simulated environment, these features could be applied to the dataset directly. The features could also be implemented separately in different datasets and compared to each other to analyse their impact on the object detections capabilities. By utilising this method for the features mentioned above, the following results could be determined. For object detection to consider most landing cases and different runways, the dataset needs to replicate real flight data and generate additional extreme landing cases. The dataset also needs to consider landings at different altitudes, which can differ at a different airport. Environmental conditions such as clouds and time of day reduce detection capabilities far from the runway, while attitude and runway appearance reduce it at close range. Runway appearance did also affect the runway at long ranges but only for darker runways.

Light Detection and Ranging (LiDAR) sensors have many different application areas, from revealing archaeological structures to aiding navigation of vehicles. However, it is challenging to interpret and fully use the vast amount of unstructured data that LiDARs collect. Automatic classification of LiDAR data would ease the utilization, whether it is for examining structures or aiding vehicles. In recent years, there have been many advances in deep learning for semantic segmentation of automotive LiDAR data, but there is less research on aerial LiDAR data. This thesis investigates the current state-of-the-art deep learning architectures, and how well they perform on LiDAR data acquired by an Unmanned Aerial Vehicle (UAV). It also investigates different training techniques for class imbalanced and limited datasets, which are common challenges for semantic segmentation networks. Lastly, this thesis investigates if pre-training can improve the performance of the models. The LiDAR scans were first projected to range images and then a fully convolutional semantic segmentation network was used. Three different training techniques were evaluated: weighted sampling, data augmentation, and grouping of classes. No improvement was observed by the weighted sampling, neither did grouping of classes have a substantial effect on the performance. Pre-training on the large public dataset SemanticKITTI resulted in a small performance improvement, but the data augmentation seemed to have the largest positive impact. The mIoU of the best model, which was trained with data augmentation, was 63.7% and it performed very well on the classes Ground, Vegetation, and Vehicle. The other classes in the UAV dataset, Person and Structure, had very little data and were challenging for most models to classify correctly. In general, the models trained on UAV data performed similarly as the state-of-the-art models trained on automotive data.

Les modèles d’apprentissage profond à large capacité ont souvent tendance à présenter de hauts écarts de généralisation lorsqu’ils sont entrainés avec une quantité limitée de données étiquetées. Dans ce cas, des réseaux de neurones très profonds et larges auront tendance à mémoriser les échantillons de données et donc ils risquent d’être vulnérables lors d’un léger décalage dans la distribution des données au moment de tester. Ce problème produit une généralisation pauvre lors de changements dans la répartition des données au moment du test. Pour surmonter ce problème, certaines méthodes basées sur la dépendance et l’indépendance de données ont été proposées. Une récente classe de méthodes efficaces pour aborder ce problème utilise plusieurs manières de contruire un nouvel échantillon d’entrainement, en mixant une paire (ou plusieurs) échantillons d’entrainement. Dans cette thèse, nous introduisons PatchUp, une régularisation de l’espace des caractéristiques au niveau des blocs dépendant des données qui opère dans l’espace caché en masquant des blocs contigus parmi les caractéristiques mappées, sélectionnés parmi une paire aléatoire d’échantillons, puis en mixant (Soft PatchUp) ou en échangeant (Hard PatchUp) les blocs contigus sélectionnés. Notre méthode de régularisation n’ajoute pas de surcharge de calcul significative au CNN pendant l’entrainement du modèle. Notre approche améliore la robustesse des modèles CNN face au problème d’intrusion du collecteur qui pourrait apparaitre dans d’autres approches de mixage telles que Mixup et CutMix. De plus, vu que nous mixons des blocs contigus de caractéristiques dans l’espace caché, qui a plus de dimensions que l’espace d’entrée, nous obtenons des échantillons plus diversifiés pour entrainer vers différentes dimensions. Nos expériences sur les ensembles de données CIFAR-10, CIFAR-100, SVHN et Tiny-ImageNet avec des architectures ResNet telles que PreActResnet18, PreActResnet34, WideResnet-28-10, ResNet101 et ResNet152 montrent que PatchUp dépasse ou égalise les performances de méthodes de régularisation pour CNN considérée comme état de l’art actuel. Nous montrons aussi que PatchUp peut fournir une meilleure généralisation pour des transformations affines d’échantillons et est plus robuste face à des attaques d’exemples contradictoires. PatchUp aide aussi les modèles CNN à produire une plus grande variété de caractéristiques dans les blocs résiduels en comparaison avec les méthodes de pointe de régularisation pour CNN telles que Mixup, Cutout, CutMix, ManifoldMixup et Puzzle Mix. Mots clés: Apprentissage en profondeur, Réseau Neuronal Convolutif, Généralisation,Régularisation, Techniques de régularisation dépendantes et indépendantes des données, Robustesse aux attaques adverses. / Large capacity deep learning models are often prone to a high generalization gap when trained with a limited amount of labeled training data. And, in this case, very deep and wide networks have a tendency to memorize the samples, and therefore they might be vulnerable under a slight distribution shift at testing time. This problem yields poor generalization for data outside of the training data distribution. To overcome this issue some data-dependent and data-independent methods have been proposed. A recent class of successful methods to address this problem uses various ways to construct a new training sample by mixing a pair (or more) of training samples. In this thesis, we introduce PatchUp, a feature-space block-level data-dependent regularization that operates in the hidden space by masking out contiguous blocks of the feature map of a random pair of samples, and then either mixes (Soft PatchUp) or swaps (Hard PatchUp) these selected contiguous blocks. Our regularization method does not incur significant computational overhead for CNNs during training. Our approach improves the robustness of CNN models against the manifold intrusion problem that may occur in other state-of-the-art mixing approaches like Mixup and CutMix. Moreover, since we are mixing the contiguous block of features in the hidden space, which has more dimensions than the input space, we obtain more diverse samples for training towards different dimensions. Our experiments on CIFAR-10, CIFAR-100, SVHN, and Tiny-ImageNet datasets using ResNet architectures including PreActResnet18, PreActResnet34, WideResnet-28-10, ResNet101, and ResNet152 models show that PatchUp improves upon, or equals, the performance of current state-of-the-art regularizers for CNNs. We also show that PatchUp can provide a better generalization to affine transformations of samples and is more robust against adversarial attacks. PatchUp also helps a CNN model to produce a wider variety of features in the residual blocks compared to other state-of-the-art regularization methods for CNNs such as Mixup, Cutout, CutMix, ManifoldMixup, and Puzzle Mix. Key words: Deep Learning, Convolutional Neural Network, Generalization, Regular-ization, Data-dependent and Data-independent Regularization Techniques, Robustness to Adversarial Attacks.

No description available.