Dr. Sergei Bikovskii Profile

Dr. Sergei Bikovskii

at ITMO Univ

SPIE Involvement:

Author

Publications (2)

This will count as one of your downloads.

You will have access to both the presentation and article (if available).

DOWNLOAD NOW

This content is available for download via your institution's subscription. To access this item, please sign in to your personal account.

Email or Username Forgot your username?

Password Forgot your password?

Show

Keep me signed in

No SPIE account? Create an account

Proceedings Article | 23 November 2024 Poster + Paper

Paired data synthesis for image restoration

Aleksei Denisov, Sergei Bykovskii, Sergei Tabunshchik, Andrei Zhdanov

Proceedings Volume 13239, 132391V (2024) https://doi.org/10.1117/12.3038148

KEYWORDS: Data modeling, Education and training, Diffusion, Lawrencium, Image quality, Image restoration

Read Abstract +

Image super-resolution is one of the key problems of image restoration. While a lot of methods focus on working with known degradation (like bicubic downsampling), in real-world use cases the degradation models are complex and unknown and therefore difficult to prepare for. To train such a model in supervised manner paired data with real-world degradations is needed. As only unpaired high- and low-resolution data is usually available due to the high cost of real paired data collection and alignment, methods that tackle the blind super-resolution task face the problem of degradation choice for training data generation. Some of existing methods provide a degradation pipeline that include noise injection, jpeg compression, downsampling with different kernels, etc. These methods may be effective in some cases but offer no mechanism of the pipeline to adapt to real-world scenario therefore lacking the performance. The other approach presented in this paper is simulating the degradation directly without trying to construct it from a predefined list. This can be done using modern generative models such as diffusion models. These models has strong generalization capabilities and are known to simulate the data distributions well. The proposed method uses a diffusion model trained for low-resolution image generation to simulate the degradations and construct paired data given high-resolution data. We compare proposed diffusion-based method with the existing paired data generation techniques and show the performance boost for it.

Proceedings Article | 23 November 2024 Paper

Neural processor microarchitecture design for CNN processing acceleration

Sergei Tabunshchik, Sergei Bykovskii, Pavel Kustarev

Proceedings Volume 13239, 132390L (2024) https://doi.org/10.1117/12.3036453

KEYWORDS: Matrices, Mixed reality, Data processing, Portability, Prototyping, Design, Computer hardware, Virtual reality, Neurons, Field programmable gate arrays

Read Abstract +

Mixed reality systems that integrate the real and digital worlds have recently gained popularity. The development of mixed reality devices in the form of wearable devices has been a significant trend in recent years. Key challenges in designing wearable mixed reality devices include reducing power consumption, enhancing productivity, and minimizing the size of the device. To address these challenges, neural processors are being incorporated into mixed reality systems. These are specialized hardware accelerators for convolutional neural network (CNN) processing. The use of neural processors is due to the fact that most environmental analysis and visualizations in mixed reality rely on CNNs. The use of these energy-efficient, high-performance processing devices enables increased comfort in using the wearable device through the high-performance processing of environmental data on a compact device.

The paper is devoted to the design of the microarchitecture of a neural processor for hardware acceleration of CNN processing based on the author's architecture of processor. The paper presents various microarchitectural solutions that can be used to accelerate CNN processing. We explore methods to optimize hardware resources and reduce time required for CNN processing. To achieve high throughput in pipelined computation, different algorithms for convolution calculations in a systolic array are examined. Based on the results of this research, we provide estimates of the characteristics of the neural processor with the proposed microarchitecture.

My Library

You currently do not have any folders to save your paper to! Create a new folder below.

Folder Name

Folder Description

View contact details

UPDATE YOUR PROFILE

Is this your profile? Update it now.

Sign into your SPIE.org account

Don’t have a profile and want one?

Create an account on SPIE.org

Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks. You are receiving this notice because your organization may not have SPIE eBooks access.*

*Shibboleth/Open Athens users─please sign in to access your institution's subscriptions.

To obtain this item, you may purchase the complete book in print or electronic format on SPIE.org.

ORGANIZATIONAL
Sign in with credentials provided by your organization.

Organizational Username

Organizational Password

Show/Hide Password

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available

Members:

Non-members: ADD TO CART

Keywords/Phrases

Search In:

Publication Years