Compressing Vision Transformers for Efficient Image Processing

Konzulens:

Bencsik Blanka

Tárgy:

Önálló laboratórium 1 - Irányítórendszerek főspecializáció, MSc Vill.
Önálló laboratórium 2 - Irányítórendszerek főspecializáció, MSc Vill.
Önálló laboratórium 1 - Vizuális informatika főspecializáció, MSc Info.
Önálló laboratórium 2 - Vizuális informatika főspecializáció, MSc Info.
Önálló laboratórium 1 - Intelligens beágyazott mecha. rendsz. szakir., MSc Mecha.
Önálló laboratórium 1 - Irányító és látórendszerek MSc. főspec.
Önálló laboratórium 1 - Vizuális informatika MSc. főspec.
Önálló laboratórium 2 - Irányító és látórendszerek MSc. főspec.
Önálló laboratórium 2 - Vizuális informatika MSc. főspec.

Hallgatói létszám:

Folytatás:

Szakdolgozat / Diplomaterv
TDK dolgozat

Leírás:

Convolutional Neural Networks (CNN) have long been the preferred choice for image processing tasks, excelling in areas like image classification, object detection, and segmentation. Their strength lies in capturing local spatial patterns through convolutional layers, facilitating hierarchical feature extraction. Conversely, Transformers, known for their computational efficiency and scalability in NLP, have recently found application in image processing as Vision Transformers. They have demonstrated remarkable performance, often matching or surpassing state-of-the-art results in various image recognition benchmarks. Therefore, Vision Transformers hold significant promise for the future of computer vision tasks.

However, it is worth noting that training and applying multi-head transformers come with a considerable computational cost. Consequently, compressing these models becomes a crucial step to enhance their practical applicability in the future.

The task for the student is to get familiar with Vision Transformers and understand their distinctions from the commonly used CNNs. Subsequently, their objective is to solve an image classification or detection task using Vision Transformers. Eventually, they are to explore and experiment with pruning techniques applicable for compression.