Hypertension – high blood pressure (BP) – is known to be a silent killer. Untreated, it can cause
severe damage to the human’s organs, mainly to the heart and kidneys [5, 6]. BP is
usually classified by using the highest – systolic – and the lowest – diastolic – pressures during one
cardiac cycle . The gold standard for measuring BP remains the oscillometric method,
which is employed in traditional arm-cuffs . This method, however, suffers from extensive
deficiencies: Discomfort leads to unreliable measures . Additionally, it only captures
the static status of the very dynamic arterial BP and thus loses important variation information,
leading to poor time resolution [2, 3, 4, 7] However, there is a strong
need for continuous beat-to-beat BP readings , as they are more reliable predictors of
aforementioned cardiovascular risks than single readings .
The goal of this master thesis is to show whether it is feasible to use a 60GHz radar device to
continuously estimate BP. Radar is chosen as it has a very small form factor and very low power
consumption – both being favorable characteristics for integrating into a wearable device. The
radar is put into an 3D-printed enclosure which is fastened to the left wrist using a velcro strap. It
is capable of extracting the skin displacement caused by the expansion of the underlying artery,
which is localized using a beamforming algorithm. The extracted skin displacement contains the
pulse waveforms which are used for extracting the BP.
In literature, mainly two methods have been used to design continuous BP devices. One is based
on Pulse-Wave-Velocity, and in that context also Pulse-Transit-Time, the other is based on Pulse-
Wave-Analysis . Since the first method depends on the usage of an electrocardiograph,
this method was not employed in this work, as the goal is to implement a stand-alone solution
which does not require additional devices. Therefore, the second method is implemented.
For that, the extracted skin displacement is split into individual pulse waveforms. Each is used
as input for a support vector machine, that decides whether it is good enough as an input for the
neural network, such that only sufficiently good waveforms are used. Then, 21 distinctive features
are extracted for the individual good waveforms. These features, together with the calibration
parameters gender, age, height and weight, are used as features for a neural network. The network
is then used to predict systolic and diastolic values.
It is expected that some correlation between the skin displacement, captured by the radar, and
the corresponding BP will become apparent, allowing for future research to further improve the
 D. Buxi, J.-M. Redout´e, and M. R. Yuce. Blood pressure estimation using pulse
transit time from bioimpedance and continuous wave radar. IEEE Transactions on
Biomedical Engineering, 64(4):917–927, 2016.
 Y. Kurylyak, F. Lamonaca, and D. Grimaldi. A neural network-based method for
continuous blood pressure estimation from a ppg signal. In 2013 IEEE International
instrumentation and measurement technology conference (I2MTC).
 M. Proenc¸a, G. Bonnier, D. Ferrario, C. Verjus, and M. Lemay. Ppg-based blood
pressure monitoring by pulse wave analysis: calibration parameters are stable for three
months. In 2019 41st Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), pages 5560–5563. IEEE, 2019.
 J. Sol`a and R. Delgado-Gonzalo. The handbook of cuffless blood pressure monitoring.
Springer. Available online at: https://link. springer. com/book/10, 1007:978–3, 2019.
 WHO. Hypertension.World Health Organization. URL: https://www.who.int/
 X. Xing, Z. Ma, M. Zhang, Y. Zhou, W. Dong, and M. Song. An unobtrusive and
calibration-free blood pressure estimation method using photoplethysmography and
biometrics. Scientific reports, 9(1):1–8, 2019.
 Y. Yoon, J. H. Cho, and G. Yoon. Non-constrained blood pressure monitoring using
ecg and ppg for personal healthcare. Journal of medical systems, 33(4):261–266, 2009.
Proper initialization of convolution kernel is crucial for a highly optimized deep learning neural network . A popular way to instantiate these kernels is random assignment of weights . It follows a gaussian distribution pattern with a mean value of 0 and standard deviation of 1. Despite being easy to implement in a neural network it has quite a few downsides like not finding the global optima or slowing the training process down. As a further improvement to random assignment Xavier Glorot et al. proposed “Xavier initialization value (Xe)”  for convolution kernels. This method follows an uniform distribution with a 0 mean and a variance of 1/n where n is the total number of input neurons. Although, training process is faster with increased convergence speed, the derivation process of Xe initialization is based on the assumption that the activation function is linear, which is not the case for popular activation functions such as Rectified Linear Unit (Relu). To mitigate this issue Kaiming He et al. proposed He initialization  targeted more toward Relu activation function. He uses a gaussian uniform distribution of 0 mean and a variance of 2/n. All of the above initialization techniques for convolution kernels are based on independent initialization of kernel weights, not taking into account the already available data of training samples. The kernel weights are trained in such a way that these randomly generated values are tried to be matched against the local pattern of the images. In every iteration, the trainer tries to minimize the error between the kernel weights and the local features, which leads further to convergence. As this is a probability event, so it takes quite a lot of iteration after which the convolution kernels can have better match with the local features. This translates into slow down of network, with a larger training time and longer convergence rate. Different methods for initializing the convolution kernel have taken this issue into account. OrthoNorm is another method that uses orthogonal matrix for kernel initialization. It can successfully be used in non-linear networks as well unlike random assignment . There is also “Layer sequence unit variance (LSUV)” method which takes the orthogonal initialization to the iterative process. It uses singular value decomposition SVD to replace the weights initiated with gaussian noise . In 2014 Tsung-HanChan et al. proposed a Principal Component Analysis (PCA) based method for convolution kernel initialization . The model gets all image patches from a feature map and initializes
the convolution kernel by calculating the principal components of image patches. This thesis aims to further improve the PCA based kernel initialization method by incorporating
ground truth GT images. GT images are already labeled and can be used to find suitable feature sets. Leveraging the dominant features from these sets and using them as convolution kernel weights, a dependency between training images and convolution kernels is created. It could theoretically decrease the training time and improve overall convergence rate . Extensive benchmarking of the proposed initialization method along with other quantitative measures needs to be taken into account
while developing the system which is also included in the scope of this thesis. To achieve the goals of the thesis work, already existing tools and libraries such as, Pytorch Lightning(
www.pytorchlightning.ai), Monai(monai.io),Weights and Biases(wandb.ai), python(www.python.org) and notable python scientific packages shall be used and re-used where possible.
The thesis will comprise the following work items:
Literature overview of improved convolution kernel initialization method
Design and formalization of the system to be developed
Overview and explanation of the algorithms used
System development including code implementation
Quantitative evaluation of the implemented system on medical image data
 Chunyu Xu and Hong Wang. Research on a convolution kernel initialization method for speeding
up the convergence of cnn. Applied Sciences, 12:633, 01 2022.
 Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional
neural networks. In F. Pereira, C.J. Burges, L. Bottou, and K.Q. Weinberger, editors,
Advances in Neural Information Processing Systems, volume 25. Curran Associates, Inc., 2012.
 Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward
neural networks. In Yee Whye Teh and D. Mike Titterington, editors, AISTATS, volume 9 of
JMLR Proceedings, pages 249–256. JMLR.org, 2010.
 Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing
human-level performance on imagenet classification, 2015.
 Andrew M. Saxe, James L. McClelland, and Surya Ganguli. Exact solutions to the nonlinear
dynamics of learning in deep linear neural networks, 2013.
 Dmytro Mishkin and Jiri Matas. All you need is a good init, 2015.
 Tsung-Han Chan, Kui Jia, Shenghua Gao, Jiwen Lu, Zinan Zeng, and Yi Ma. PCANet: A
simple deep learning baseline for image classification? IEEE Transactions on Image Processing,
24(12):5017–5032, dec 2015.
Current deep-learning-based classification methods require large amounts of data for training, and in certain scenarios such as in the surveillance imaging there is only a limited amount of data. The aim of the research is to generate new training images of vehicles with the same characteristics as the training data but from novel view points and investigate its suitability for fine-grained classification of vehicles.
Generative models such as generative adversarial networks (GANs)  allow for customization of images. However, adjusting the perspective through methods such as conditional GANs for unsupervised image-to-image translation has proven to be particularly difficult . Methods such as StyleGANs  or neural radiance fields (NeRFs)  are relevant approaches to generate images with different styles and perspectives.
StyleGAN is an extension to the GAN architecture that proposes changes to the generator model such as the introduction of a mapping network. The mapping network generates intermediate latent codes which are transformed into styles that is integrated at each point in the generator network. It also includes a progressive growing approach for training generator models capable of synthesizing very large high-quality images.
NeRF can generate novel views of complex 3D scenes based on a partial set of 2D images. It is trained to directly map from spatial location and viewing direction (5D input) to opacity and color, using volume rendering  to render new views.
The thesis consists of the following milestones:
- Literature review on the state-of-the-art approaches for GAN- and neural radiance fields-based
- Adoption of existing GAN- and neural radiance fields-based image synthesis methods to generate
car images using different styles and camera poses 
- Experimental evaluation and comparison of different image synthesis methods
- Investigate the suitability of the generated images for fine-grained vehicle classification using
different classification methods , 
The implementation will be done in Python.
 Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, “Generative Adversarial Networks ”, in NIPS, 2014
 Tero Karras, Samuli Laine, Timo Aila, “A Style-Based Generator Architecture for Generative Adversarial Networks ”, in proceedings of the IEEE/CVF Conference on CVPR, 2019
 Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng, “NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis ”, in ECCV 2020
 Robert A. Drebin, Loren Carpenter, Pat Hanrahan, “Volume Rendering ”, in Proceedings of SIGGRAPH 1988
 Jiatao Gu, Lingjie Liu, Peng Wang, Christian Theobalt, “StyleNeRF: A Style-based 3D-Aware Generator for High-resolution Image Synthesis ”, in ICLR 2022
 Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo, “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows ”, in IEEE/CVF conference on ICCV, 2021
 Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, “Deep residual learning for image recognition ”, Proceedings of the IEEE conference on CVPR, 2016