PhD low-complexity learning-based video compression

Job title:

PhD low-complexity learning-based video compression

Company:

Orange

Job description

about the roleYour role will be to carry out a PhD thesis on low-complexity learning-based video compression.

• Global context and thesis problem statement

The volume of video content exchanged online is constantly growing, with new video formats emerging regularly. Compression tools are evolving to handle this content efficiently, aiming for lower bitrates while maintaining good quality.Standardization groups like ISO/MPEG and ITU-T have established successive standards (AVC, HEVC, VVC) over decades, refining the “conventional” compression paradigm. A new standard is expected by 2030.Recently, neural network-based compression systems have emerged. They replace the traditional encoder-decoder (codec) with a generic neural network (autoencoder) trained on vast video datasets to handle diverse content and formats. These autoencoders rival the best conventional codecs in performance. However, their decoding complexity remains a significant limitation.One approach to address this consists in replacing the generic autoencoder with a neural network specifically (over-)fitted to the image being compressed. This involves training a small neural network during each image’s compression, which is then used for decoding. Orange and other actors are developing a codec based on this overfitting approach called Cool-chic. Cool-chic achieves image coding performance comparable to autoencoders while being 1000 times less complex. Initial efforts are underway to extend Cool-chic for video compression.

• Scientific Goals – Results and technical challenges

The thesis aims to improve the exploitation of temporal redundancies in a neural video codec based on overfitting.Current overfitting-based codecs struggle to efficiently leverage temporal redundancies. They rely on motion analysis and temporal prediction to identify redundant areas in successive video frames. One research direction is to enhance this process through improved motion analysis, leading to better temporal prediction. Additionally, new neural network architectures offer the potential to remove temporal redundancies without explicit motion analysis.Overfitting-based codecs have a unique decoder for each image, which needs to be transmitted alongside the compressed signal. This auxiliary signal differs from typical image data, as it consists of neural network parameters. The thesis will explore ways to improve the transmission of this signal, potentially by sharing elements between successive neural networks within a video.Maintaining a low complexity is crucial to ensure limited energy consumption across various target devices.about you

• Scientific and technical skills for this position

Knowledge on signal processing and deep learningInterest for video processingProgramming: Python, C++, bash etc.Experience using deep learning frameworks: PyTorch, Tensorflow, jax

• Education

Master’s degree or engineering degree

• Previous professional experiences

Experiences in the field of image processing and/or machine learningadditional informationThis thesis aims to enhance overfitting-based neural codecs by leveraging redundancies, advancing the field of AI-driven video coding.Initially, the thesis will map out various tools for exploiting information available at the decoder. This involves a comprehensive review of motion compensation mechanisms. Additionally, it will analyze techniques like residual coding and conditional coding. A broader literature review will cover signal processing and machine learning techniques applied to image and video compression.Based on this research, a technical development path will be proposed to improve an existing overfitting-based neural video codec (Cool-chic) by optimizing redundancy exploitation.Development will be collaborative with team researchers, involving:

• Open-source code development for AI-based compression.
• Contribution to the definition of future video compression standards.
• Authorship of scientific articles and presentations at international conferences.
• Patent drafting.
• Collaboration with other academic and industrial labs partnered with Orange.

References

• Cool-chic, A low-complexity AI-based codec:
• https://orange-opensource.github.io/Cool-Chic/
• https://arxiv.org/abs/2212.05458

departmentOrange Innovation brings together the research and innovation activities and expertise of the Group’s entities and countries. We work every day to ensure that Orange is recognized as an innovative operator by its customers and we create value for the Group and the Brand in each of our projects. With 720 researchers, thousands of marketers, developers, designers and data analysts, it is the expertise of our 6,000 employees that fuels this ambition every day.Orange Innovation anticipates technological breakthroughs and supports the Group’s countries and entities in making the best technological choices to meet the needs of our consumer and business customers.You will be integrated into a multidisciplinary research team at the forefront of innovation and expertise in video compression, as well as 3D video and audio signal processing. The team has long played a key role in the international standardization process (ISO/MPEG and ITU-T), which adds an applied and dynamic aspect to this doctoral study.The thesis addresses a topic of significant stakes, as it is focused on a key industrial challenge for Orange and many other global players. Thus, the results of this thesis will be directly applicable, as neural video compression is an important technical concern looking towards 2030.contractThesis

Expected salary

Location

Caen, Calvados

Job date

Tue, 04 Mar 2025 23:42:34 GMT

To help us track our recruitment effort, please indicate in your email/cover letter where (vacanciesin.eu) you saw this job posting.