JPEG-AI learned image compression on RISC-V architectures (master)
Description: The Learned Image Compression (LIC) paradigm [1] is meant to replace traditional image compression methods with autoencoder-based models that can be trained end-to-end at image compression. First, the image is projected into a low-dimensional latent space by the encoder. This representation is then quantized and entropy-coded into a binary bitstream that is decoded at the receiver side, recovering an approximation of the original image. Advances in LIC have spurred the interest of the image compression industry and the Joint Picture Experts Group (JPEG) is currently standardizing a LIC-based image compression format known as JPEG-AI. While LIC based image compression may exceed traditional methods in terms of compression efficiency, that usually comes at the cost of high computational encoding and decoding complexity due to the deep stacks of convolutional layers that are involved in the encoding and decoding process. Realtime-capable JPEG-AI implementations have been demonstrated using specialized parallel accelerators such as GPUs and FPGAs, yet little attention has been devoted to CPU-based implementations and especially to RISC-V based implementations. RISC-V (Reduced Instruction Set Computer-Five) has recently attracted significant attention as a promising architecture due to its open-source Instruction Set Architecture (ISA), as well as its flexibility, scalability, and high degree of customization. It offers a compelling balance between performance and power efficiency while enabling the exploitation of multiple levels of parallelism. This thesis aims at exploring RISC-V based JPEG-AI solutions, leveraging RISC-V Vector SIMD extensions or specialized edge AI solutions. In particular, the work focuses on optimizing linear algebra operations, such as convolutions, required for real-time JPEG-AI using both automatic compiler vectorization and vector intrinsic instruction techniques.
Contacts: Attilio Fiandrotti (attilio.fiandrotti@unito.it) Gianluca Dalmasso (gianluca.dalmasso@unito.it) Iacopo Colonnelli (iacopo.colonnelli@unito.it) Robert Birke (robert.birke@unito.it)
Internship in Internal Startup of Thales Alenia Space in Torino (master)
Description: Thales Alenia Space in Torino is a key player of space exploration, currently involved in the development of many programs for Earth Orbit (EUCLID, Axiom Hotel) the Moon (Lunar Gateway, Argonaut Lunar Lander, Moon Village), Mars (ExoMars Rosalind Franklin) and beyond. As part of its Open Innovation, Thales Alenia Space increased its Space Business Catalyst footprint by implementing an antenna in Torino to cover activities of the domain Exploration and Science. Incubated in the Space Business Catalyst in Torino, the first Italian, internal startup, Sparks, to find new business in non-core areas. The team is therefore recruiting an intern to support them in these activities. Join us on its innovative mission to democratize space exploration! We are developing a unique Space-themed digital crowdfunding platform, mixing space education and mobile casual video games in an edutaining and playful manner. Our project lies at the crossroads of blockchain technology and edutainment, aiming to build a unified, cooperative, and expansive Space community globally. If you share our passion for technology, education, and space exploration, this is the perfect opportunity for you!
Contacts: Luca Simonini (luca.simonini@thalesaleniaspace.com) Agata Soccini (agatamarta.soccini@unito.it)
Robust AI-based Video Compression (master/bachelor)
Description: Nowadays over 60% of the whole Internet traffic is represented by video communications, therefore efficient video compression techniques able to relieve the pressure on the network infrastructure are in demand. Over the past years, the deep learning paradigm has enjoyed tremendous success in a number of fields, including image and video compression [2]. AI-based video codecs built atop of deep neural architectures nowadays achieve compression efficiency not far from standardized video codecs [3] such as MPEG H.264/AVC (DVBT) or H.265/HEVC (DVBT2). However, existing research in the field focuses mainly on improving the compression efficiency, whereas the issues related to the transmission of the AI-compressed video over unreliable channels (e.g., wireless channels affected by noise or packet networks affected by router congestion) have been explored little. This thesis proposal lies at the boundary between AI, deep learning and image compression and aims at investigating the effect of data losses over the transmitted video as a first step, and devising robust learnable encoding algorithms as a potential further step. The ideal profile includes knowledge of deep learning techniques on one side and familiarity with video coding methods on the other and familiarity with the Python language.
Contacts: Attilio Fiandrotti (attilio.fiandrotti@unito.it) Valerio Bioglio (valerio.bioglio@unito.it)
Multi-target stain normalization for histopathology (master/bachelor) [pdf]
Description: Develop and implement novel techniques for stain normalization in histological images for deep learning applications.