Post-Doctoral Research Visit F/M Universality for Deriving Low-Complexity INR for Visual Compression Models in Rennes

Context

This postdoctoral position is part of the joint research activities between Inria and InterDigital conducted within Nisk.AI, a joint research project focusing on efficient and frugal compression of new media through artificial intelligence and neural network-based coding. Nisk.AI brings together Inria research teams and InterDigital to investigate new representations and coding technologies that can improve the efficiency of visual data compression while addressing the computational challenges associated with AI-based approaches. The project provides a strong academic-industrial research environment, combining Inria's expertise in information theory, signal and image processing, and machine learning with InterDigital's expertise in media technologies and standardization.

Assignment

One of the factors responsible for the growing popularity of Implicit Neural Representation (INR) models for image compression is their relatively low decoding time and their compression efficiency [1]. INR-based image compression introduces a new paradigm in which each image is represented by a function that can be transmitted and subsequently used to reconstruct the image [2]. Once this function has been learned at the encoder and transmitted, the image can be reconstructed from the learned representation. However, one of the main challenges of this paradigm is the long image encoding time. This is because learning the personalized function requires "overfitting" a separate model to each individual image or video, which can be computationally expensive and time-consuming [3]. This remains one of the main limitations to the practical deployment of INR-based compression methods [4].

Main activities

In this project, we aim to investigate whether common knowledge can be extracted from the functions learned for individual images, thereby reducing the need to learn each function from scratch. More specifically, we seek to answer the following question: Can common knowledge be extracted from image-specific functions in order to reduce the training required for each new image? To address this question, we will analyze learned INR functions and investigate whether some form of universality across images can be identified. Based on these insights, we will explore strategies for reducing the training complexity of INR-based architectures and, consequently, their image encoding time. The expected outcome is to develop INR-based compression models that retain the advantages of implicit representations while significantly reducing their computational cost at the encoder.

Objectives

The main objectives of this project are:

• To investigate how INR-based models relate to conventional autoencoder architectures for image compression, and whether they exhibit common characteristics;
• To determine whether a subset of internal features or representations can be identified across different images;
• To exploit the universality of these features to develop lower-complexity INR-based image compression models with faster encoding times.

References

• [1] T. Ladune, P. Philippe, F. Henry, G. Clare, and T. Leguay, "Cool-chic: Coordinate-based low complexity hierarchical image codec," in International Conference on Computer Vision (ICCV), 2023, pp. 13 515-13 522.
• [2] E. Dupont, A. Goliński, M. Alizadeh, Y. W. Teh, and A. Doucet, "COIN: Compression with implicit neural representations," 2021. [Online]. Available: https://arxiv.org/abs/2103.03123
• [3] E. Dupont, H. Loya, M. Alizadeh, A. Golinski, Y. W. Teh, and A. Doucet, "COIN++: Neural compression across modalities," Transactions on Machine Learning Research, 2022. [Online]. Available: https://openreview.net/forum?id=NXB0rEM2Tq
• [4] M. Gwilliam, R. Zhang, N. Padmanabhan, H. Du, and A. Shrivastava, "How to design and train your implicit neural representation for video compression," in Winter Conference on Applications of Computer Vision (WACV), 2026.

#J-18808-Ljbffr