Synthetic image generation for anonymisation
Your tasks
The aim of the project is to test the state-of-the-art image generation networks called generative diffusion models (GDMs) and compare their performance to generative adversarial networks (GANs). The methods will be tested with natural image datasets (MNIST, CIFAR10) as well as digital pathology datasets. You will have the opportunity to have weekly supervision by the experts of the field, get hands-on experience with deep learning methods, and enjoy the flexible working environment of academia. You will work at the center for medical imaging and visualization (CMIV) located at the Linköping’s university hospital.
Your profile
We are looking for 1-2 students with background in image processing and machine learning. You have a great interest in image generation and artificial intelligence, and in the applications within medicine.
Information
Research group: Computer graphics and image processing
Contact person: Milda Pocevičiūtė
Location: Center for medical imaging and visualization, CMIV
Main field of studies: Medieteknik
Keywords: machine learning, deep learning, image generation, medical imaging
Level: Master
Data augmentation via synthetic image generation
Digital pathology patch (leftmost) transformed by a generative diffusion model.
Training robust artificial intelligence (AI) algorithms require immense amounts of high-quality data. However, collecting large and diverse medical data often is unfeasible due to the high costs associated with the data collection as well as the fact that some types of diseases are very rare. This results in a serious challenge for development of AI solutions for medical applications. A potential solution for alleviating the problem is to make extensive use of data augmentation techniques, boosting the diversity of available data for training. While conventional methods for data augmentation could improve performance, they are also limited. With the breakthrough in synthetic image generation by means of generative deep learning, there is a large potential in using these for advanced strategies for data augmentation.
Your tasks
The aim of the project is to evaluate different data augmentation strategies and assess if they results in improved performance and robustness of the AI algorithms. You will explore data augmentation methods based on various classical image processing (such as rotation, zooming, colour jittering, etc), the state-of-the-art image generation networks called generative diffusion models (GDMs), and the combination of both. The methods will be tested with natural image datasets (MNIST, CIFAR10) as well as digital pathology datasets. You will have the opportunity to have weekly supervision by the experts of the field, get hands-on experience with deep learning methods, and enjoy the flexible work- ing environment of academia. You will work at the center for medical imaging and visualization (CMIV) located at the Linköping’s university hospital.
Your profile
We are looking for 1-2 students with background in image processing and ma- chine learning. You have a great interest in image generation and artificial intelligence, and in the applications within medicine.
Information
Research group: Computer graphics and image processing
Contact person: Milda Pocevičiūtė
Location: Center for medical imaging and visualization, CMIV
Main field of studies: Medieteknik
Keywords: machine learning, deep learning, image generation, medical imaging
Level: Master
Visualization and importance sampling in deep learning
Machine learning has made great progress over the last decade, specifically within deep learning where artificial neural networks can model complex relationships between different types of data. When a neural network is trained, each data sample is usually given the same weight, or importance. However, in most cases this is not optimal. Many samples are simple, and do not contribute to the optimization, while others could have a negative impact on the model. Thus, with a strategy for measuring the importance of individual data samples, the training can be made more efficient and provide a better model. Also, information about sample importance can be visualized and used to provide an understanding for how data impacts the model training.
Your tasks
You will work in close collaboration with researchers within the field. You will investigate different strategies for importance sampling and hard example mining, and use different types of visualization for analyzing the results. The aim of the project is two-fold. First, the investigated techniques for importance sampling can improve model performance and decrease training times. Second, by analyzing and visualizing the sample importance across the dataset, an increased understanding can be formed regarding how the dataset should be composed. Such information is an essential building block in a human-in-the-loop machine learning pipeline focusing on training data, where the information can be used to improve the data in an iterative manner.
Your profile
We are looking for 1-2 students with background in visualization and machine learning. You have a great interest in artificial intelligence and the possibilities for understanding and improving deep learning by means of data analysis and visualization.
Information
Research group: Computer graphics and image processing
Contact person: Gabriel Eilertsen
Location: The division for Media and Information Technology, Campus Norrköping
Main field of studies: Medieteknik
Keywords: machine learning, deep learning, data analysis, information visualization
Level: Master
Rendering with non-parametric data-driven BRDF model
Photo-realistic rendering requires accurate modeling of the appearance of real-world materials using the bidirectional reflectance distribution function (BRDF). There are various ways to model BRDFs, and in practice due to their compact and flexible form, analytic BRDF models are often employed to estimate the surface properties. However, these models despite being efficient for rendering, are not very realistic. Measured BRDFs on the other hand can accurately model a realistic appearance, but they are often computationally expensive and consume significantly more memory, which makes them impractical for real-world applications. It has been shown, however, with sparse modeling of measured BRDFs, a non-parametric model can be defined that reduces the dimensionality of the BRDF, and therefore the rendering cost. Sparse modeling enables rendering speeds competitive with analytical models while admitting realistic modeling of BRDFs.
Your tasks
You will explore how non-parametric sparse BRDF modeling can be utilized for realistic rendering. You will modify an existing ray tracer such as PBRT/Mitsuba or write your own ray tracer to employ the non-parametric BRDF model and analyze the capability of this model for fast and realistic rendering. The source code and required tools for sparse BRDF modeling is available. An analysis of how the parameters of sparse modeling affect the quality and efficiency of rendering is required as well.
Your profile
We are looking for 1-2 students with a background in machine learning and computer graphics.
Information
Research group: Computer graphics and image processing
Contact person: Saghi Hajisharif and Ehsan Miandji
Location: The division for Media and Information Technology, Campus Norrköping
Main field of studies: Medieteknik
Keywords: machine learning, rendering, BRDF, sparse representations
Level: Master
Neural network feature visualization
Feature visualization answers questions about what a neural network — or parts of a network — are looking for by generating examples. If we want to understand individual features, we can search for examples where they have high values as exemplified in the images above. However, they can be hard to interpret and the correctness of current implementations are questionable due to a lack of ground truth comparisons. Furthermore, the initial implementation for the technique is based on the outdated Tensorflow 1, which does not even run anymore.
Your tasks
Investigate different techniques for generating feature visualizations. For example, by optimizing for a given set of values instead of only the high values. Also try to create verifiable visualizations that can serve as a basis for Tensorflow 2 and/or PyTorch implementations.
Your profile
We are looking for 1 student with an interest in machine learning and visualization.
Information
Research group: Computer graphics and image processing
Contact person: Daniel Jönsson
Location: The division for Media and Information Technology, Campus Norrköping
Main field of studies: Medieteknik
Keywords: machine learning, visualization
Level: Master
A learning-based video compression with sparse representation and entropy coding
Video streams constitute a large part of the daily internet traffic. A one hour long video at 4K resolution and 25 frames per second requires about 2TB of storage if no compression is applied. As a result, it is of utmost need to find solutions to intelligently transfer/use such large amounts of data. Modern video codecs have enabled the streaming of video data over the internet, in real-time, e.g., in a video call, or as demanded, such as YouTube video streaming or Netflix movies. In recent years, there have been some attempts to the standardization of machine learning approaches in video codecs such as MPEG video coding for machine (VCM) standards for machine-to-machine (M2M) or machine-to- human (M2H) communications, as well as JPEG AI, and JVET Neural Network Video Coding (NNVC). This project aims to employ an unsupervised machine learning approach for encoding and decoding a video using sparse representations and applying fast and accurate quantization and entropy coding on the resulting sparse coefficients.
Your tasks
Explore using machine learning methods to develop a codec for video stream- ing. The codec consists of both an encoder and a decoder. You will use an unsupervised machine learning method, named AMDE, to learn a sparse representation of the dataset from a training set. The video frames are then transformed into sparse coefficients which are then quantized and further compressed us- ing an entropy coding algorithm such as Huffman coding. You will carry out an analysis of the quality of the codec in terms of compression efficiency and encoding latency in comparison with state-of-the-art video codec approaches.
Your profile
We are looking for 1 student with an interest in machine learning, image processing, and computer graphics.
Information
Research group: Computer graphics and image processing
Contact person: Saghi Hajisharif and Ehsan Miandji
Location: The division for Media and Information Technology, Campus Norrköping
Main field of studies: Medieteknik
Keywords: machine learning, image processing, compression, rendering
Level: Master