Image representation is a key component in visual recognition systems. In visual recognition problem, the solution or the model should be able to learn and infer the quality of certain visual semantics in the image. Therefore, it is important for the model to represent the input image in a way that the semantics of interest can be inferred easily and reliably. This thesis is written in the form of a compilation of publications and tries to look into the Convolutional Networks (CovnNets) representation in visual recognition problems from an empirical perspective. Convolutional Network is a special class of Neural Networks with a hierarchical structure where every layer’s output (except for the last layer) will be the input of another one. It was shown that ConvNets are powerful tools to learn a generic representation of an image. In this body of work, we first showed that this is indeed the case and ConvNet representation with a simple classifier can outperform highly-tuned pipelines based on hand-crafted features. To be precise, we first trained a ConvNet on a large dataset, then for every image in another task with a small dataset, we feedforward the image to the ConvNet and take the ConvNets activation on a certain layer as the image representation. Transferring the knowledge from the large dataset (source task) to the small dataset (target task) proved to be effective and outperformed baselines on a variety of tasks in visual recognition. We also evaluated the presence of spatial visual semantics in ConvNet representation and observed that ConvNet retains significant spatial information despite the fact that it has never been explicitly trained to preserve low-level semantics. We then tried to investigate the factors that affect the transferability of these representations. We studied various factors on a diverse set of visual recognition tasks and found a consistent correlation between the effect of those factors and the similarity of the target task to the source task. This intuition alongside the experimental results provides a guideline to improve the performance of visual recognition tasks using ConvNet features. Finally, we addressed the task of visual instance retrieval specifically as an example of how these simple intuitions can increase the performance of the target task massively. / <p>QC 20161209</p>

With the emergence of Convolutional Neural Network (CNN) models, precision of image classification tasks has been improved significantly over these years. Regional CNN (RCNN) model is proposed to solve object detection tasks with a combination of Region Proposal Network and CNN. This model improves the detection accuracy but suffer from slow inference speed because of its multi-stage structure. The Single Stage Detection (SSD) network is later proposed to further improve the object detection benchmark in terms of accuracy and speed. However, SSD model still suffers from high miss rate on small targets since datasets are usually dominated by medium and large sized objects, which don’t share the same features with small ones. On the other hand, geometric analysis on dataset images can provide additional information before model training. In this thesis, we propose several SSD-based models with adjusted parameters on feature extraction layers by using geometric analysis on KITTI and Caltech Pedestrian datasets. This analysis extends SSD’s capability on small objects detection. To further improve detection accuracy, we propose a two-stream network, which uses one stream to detect medium to large objects, and another stream specifically for small objects. This two-stream model achieves competitive performance comparing to other algorithms on KITTI and Caltech Pedestrian benchmark. Those results are shown and analysed in this thesis as well.

Autonomous vehicles have many application possibilities within many different fields like rescue missions, exploring foreign environments or unmanned vehicles etc. For such system to navigate in a safe manner, high requirements of reliability and security must be fulfilled. This master's thesis explores the possibility to use the machine learning algorithm convolutional network on a robotic platform for autonomous path following. The only input to predict the steering signal is a monochromatic image taken by a camera mounted on the robotic car pointing in the steering direction. The convolutional network will learn from demonstrations in a supervised manner. In this thesis three different preprocessing options are evaluated. The evaluation is based on the quadratic error and the number of correctly predicted classes. The results show that the convolutional network has no problem of learning a correct behaviour and scores good result when evaluated on similar data that it has been trained on. The results also show that the preprocessing options are not enough to ensure that the system is environment dependent.

<div>Sorghum, which is grown internationally as a cereal crop that is robust to heat, drought, and disease, has numerous applications for food, forage, and biofuels. When monitoring the growth stages of sorghum, or phenotyping specific traits for plant breeding, it is important to identify and monitor the panicles in the field due to their impact relative to grain production. Several studies have focused on detecting panicles based on data acquired by RGB and multispectral remote sensing technologies. However, few experiments have included hyperspectral data because of its high dimensionality and computational requirements, even though the data provide abundant spectral information. Relative to analysis approaches, machine learning, and specifically deep learning models have the potential of accommodating the complexity of these data. In order to detect panicles in the field with different physical characteristics, such as colors and shapes, very high spectral and spatial resolution hyperspectral data were collected with a wheeled-based platform, processed, and analyzed with multiple extensions of the VGG-16 Fully Convolutional Network (FCN) semantic segmentation model.</div><div><br></div><div>In order to have correct positioning, orthorectification experiments were also conducted in the study to obtain the proper positioning of the image data acquired by the pushbroom hyperspectral camera at near range. The scale of the DSM derived from LiDAR that was used for orthorectification of the hyperspectral data was determined to be a critical issue, and the application of the Savitzky-Golay filter to the original DSM data was shown to contribute to the improved quality of the orthorectified imagery.</div><div><br></div><div>Three tuned versions of the VGG-16 FCN Deep Learning architecture were modified to accommodate the hyperspectral data: PCA&FCN, 2D-FCN, and 3D-FCN. It was concluded that all the three models can detect the late season panicles included in this study, but the end-to-end models performed better in terms of precision, recall, and the F-score metrics . Future work should focus on improving annotation strategies and the model architecture to detect different panicle varieties and to separate overlapping panicles based on an adequate quantities of training data acquired during the flowering stage.</div>

Due to hardware limitations, 3D sensors like LiDAR often produce sparse and noisy point clouds. Point cloud upsampling is the task of converting such point clouds into dense and clean ones. This thesis tackles the problem of point cloud upsampling using deep neural networks. The effectiveness of a point cloud upsampling neural network heavily relies on the upsampling module and the feature extractor used therein. In this thesis, I propose a novel point upsampling module, called NodeShuffle. NodeShuffle leverages Graph Convolutional Networks (GCNs) to better encode local point information from point neighborhoods. NodeShuffle is versatile and can be incorporated into any point cloud upsampling pipeline. Extensive experiments show how NodeShuffle consistently improves the performance of previous upsampling methods. I also propose a new GCN-based multi-scale feature extractor, called Inception DenseGCN. By aggregating features at multiple scales, Inception DenseGCN learns a hierarchical feature representation and enables further performance gains. I combine Inception DenseGCN with NodeShuffle into the proposed point cloud upsampling network called PU-GCN. PU-GCN sets new state-of-art performance with much fewer parameters and more efficient inference.

Fuel performance codes are used to demonstrate with confidencethat nuclear fuel rods will sustain normal operation and transientevents without being damaged. However, the execution time of a typ-ical fuel rod simulation ranges from tens of seconds to minutes which can be impractical in certain applications. In the scope of this work,at least two such applications are identified; code-calibration and fuelcore evaluations. In both of these cases, possible improvements can be obtainedby creating neural network surrogate models. For code calibration,a Deep Neural Network is enough since calibration is performed onmodel constants. But for full-core evaluations, a surrogate model mustbe able to predict a time-dependent target as a function of a time-dependent input. In this work, Temporal Convolutional Networks are investigated for the second application. In both applications, targetdata are generated with a Cladding Oxidation Model. The result of the study shows that both models succeeded in their respective tasks with good performance metrics. However, furtherwork is needed to increase the number of input and target variablesthat the Deep Neural Network can handle, verify the flexibility ofinput data files for the TCN, try out the TCN on a real code, and combine the two models and achieve a broader set of use-cases. / <p>Kursnamn: Fördjupande projektarbete i energisystem</p><p>Kurskod: 1FA394</p>

This thesis deals with rotation estimation of 3D model of human jaw. It describes and compares methods for direct analysis od 3D models as well as method to analyze model using rasterization. To evaluate perfomance of proposed method, a metric that computes number of cases when prediction was less than 30° from ground truth is used. Proposed method that uses rasterization, takes  three x-ray views of model as an input and processes it with convolutional network. It achieves best preformance, 99% with described metric. Method to directly analyze polygonal model as a sequence uses attention mechanism to do so and was inspired by transformer architecture. A special pooling function was proposed for this network that decreases memory requirements of the network. This method achieves 88%, but does not use rasterization and can process polygonal model directly. It is not as good as rasterization method with x-ray display, byt it is better than rasterization method with model not rendered as x-ray.  The last method uses graph representation of mesh. Graph network had problems with overfitting, that is why it did not get good results and I think this method is not very suitable for analyzing plygonal model.

No description available.

In tropical and subtropical countries with bad infrastructure there exists diseases which are often neglected and untreated. Some of these diseases are caused by parasitic intestinal worms which most often affect children severely. The worms spread through parasite eggs in human stool that end up in arable soil and drinking water. Over one billion people are infected with these worms, but medication is available. The problem is the ineffective diagnostic method hindering the medication to be distributed effectively. In the process of designing an automated microscope for increased effectiveness the solution for marking out the stool sample on the microscope slide is important for decreasing the time of diagnosis. This study examined the active contour model and four different semantic segmentation networks for the purpose of delineating the stool sample from the other parts of the microscope slide. The Intersection-over-Union (IoU) measurement was used to measure the performance of the models. Both active contour and the networks increased the IoU compared to the current implementation. The best model was the FCN-32 network which is a fully convolutional network created for semantic segmentation tasks. This network had an IoU of 95.2%, a large increase compared to the current method which received an IoU of 77%. The FCN-32 network showed great potential of decreasing the scanning time while still keeping precision of the diagnosis.

No description available.