Computer Engineering MA, Multidimensional Visual Representation and Compression, 6 credits

Aim

The course introduces theories and practical methods for the representation and compression of multidimensional visual data. It integrates fundamental concepts from information theory, source coding, and signal processing with state-of-the-art techniques in visual media, including images, video, 3D models, and light fields.

Course objectives

Upon completion of the course the student should be able to:

• Describe how 2D, video, point cloud, and light field representations are employed in visual media, focusing on efficiency, quality, and immersiveness.
• Apply various redundancy reducing approaches alongside advanced lossless coding approaches to establish core compression principles.
• Synthesize fundamental information theory concepts to develop strategies that optimize data compression across diverse formats.
• Optimize quantization and rate-distortion performance through advanced techniques that balance compression efficiency with quality.
• Compare state-of-the-art methods and deep learning-based architectures with conventional paradigms to explore modern advancements in codec design.
• Evaluate visual representation and compression techniques using objective and perceptual metrics for comprehensive quality assessment.

Content

Multidimensional Visual Data Representation: Image, video, 3D point clouds, light fields, 3D Gaussian Splats, NERFs.

Information Theory: entropy, conditional entropy, joint entropy, mutual information, energy compactness.

Prediction and transformation: linear prediction, prediction error, delta encoding, frequency transformations, sub-band, discrete cosine transform, wavelet transform, Karhunen-Loève transform.

Lossless coding: Entropy codes, fixed and variable length coding, dictionary methods, hyperprior modeling, context-adaptive compression, run-length encoding.

Lossy coding: Sub-sampling, uniform/non-uniform/adaptive, scalar/vector quantization, quantization noise, noise shaping, rate-distortion optimization, Lagrange multipliers.

Traditional vs. Deep Learning-Driven Compression: Image and video codec evolution, VAEs, GANs, Vision Transformers, neural codecs, hybrid approaches.

Quality Assessment Metrics: Objective and perceptual evaluation, MSE, PSNR, SSIM, VMAF, JND, single- and double-stimulus methods, pairwise comparison, Mean Opinion Score (MOS), and Visual Information Fidelity (VIF).

Engaging in hands-on coding exercises, experimenting with real-world datasets, studying state-of-the-art representation formats and coding approaches, considering ethical implications.

Entry requirements

Computer Engineering or Electrical Engineering, 60 credits, including one course in object-oriented programming.

Mathematical subjects, 30 credits, including probability theory and statistics, and linear algebra.

Computer Engineering MA, Signal and Image Processing, 6 credits.

Selection rules and procedures

The selection process is in accordance with the Higher Education Ordinance and the local order of admission.

Examination form

L101: Laboratory work, 1.5 Credits
Grade scale: Two-grade scale

T101: Written exam, 4.5 Credits
Grade scale: Seven-grade scale, A-F o Fx

Link to grading criteria: https://www.miun.se/gradingcriteria.

The examiner has the right to offer alternative examination arrangements to students who have been granted the right to special support by Mid Sweden University’s disabilities adviser.

Examination restrictions

Students registered on this version of the syllabus have the right to be examined 3 times within 1 year according to specified examination forms. After that, the examination form applies according to the most recent version of the syllabus.

Grading system

Seven-grade scale, A-F o Fx