An efficient implementation of 2D convolution in CNN

Jing, Chang; Sha, Jian

doi:10.1587/elex.13.20161134

Cited by 34 publications

(16 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…где k i,j -фильтры свертки, b i -значения смещения, F act -функция активации [2]. Методы реализации двухмерной свертки на ПЛИС описаны в [16]. Слой субдискретизации производит сокращение изображения в 2 раза (по горизонтали и вертикали) [Ошибка!…”

Section: рис 2 структура сверточной нейросети (1 канал)unclassified

Using Fast Prototyping Facilities for Implementation of a Convolution Neural Network on a Fpga

Bakhchevnikov¹,

Derkachev²,

Bakumenko³

2020

IZVESTIYA SFedU. ENGINEERING SCIENCES

View full text Add to dashboard Cite

Исследования в области искусственного интеллекта ведутся с возрастающим интересом с каждым годом. Области применения искусственного интеллекта довольно обширны: автоматизация, анализ большого объема данных, технологии умного дома, машинное зрение и т.д. Технологии искусственного интеллекта базируются на использовании искусственных нейронных сетей, имеющие в своей основе принципы нервной си стемы животных. При этом актуальным вопросом является реализация искусственных нейронных сетей на различных программно-аппаратных платформах: программируемых логических интегральных схемах (ПЛИС) типа FPGA (Field Programmable Gate Array), на интегральных схемах специального назначения (Application-Specific Integrated Circuit, ASIC), GPU, CPU и т.д. ПЛИС наилучшим образом проявляют себя в маломощных мобильных системах. ASIC демонстрируют наибольшую производительность с недоста тком: высокая цена разработки. Проблема быстрого прототипирования проектов, основанных на использовании искусственных нейронных сетей, для ПЛИС привычными методами (c помощью HDL-языков, HDL-кодеров, графического программирования) заключается в том, что либо такой проект сложен и длителен в отладке (HDL-языки), либо не оптимален получающийся код (HDL-кодеры), либо высока длительность разработки проекта и сложность реконфигурации нейронной сети (графическое программирование). Поэтому в рамках данной работы рассматривается эффективный метод проектирования полносвязных и сверточноых нейронных сетей для их реализации на ПЛИС использованием пакета Xilinx System Generator for DSP и Matlab/Simulink. Генерируемые таким образом искусственные нейросети легко реконфигурируемы и позволяют решать следующие задачи: распознавание изображений, оптимальная фильтрация (например, для задач подповерхностной радиолокации). Искусственный интеллект; искусственные нейронные сети; реализация на ПЛИС; сверточная нейросеть; метод проектирования.

show abstract

Section: рис 2 структура сверточной нейросети (1 канал)unclassified

Using Fast Prototyping Facilities for Implementation of a Convolution Neural Network on a Fpga

Bakhchevnikov¹,

Derkachev²,

Bakumenko³

2020

IZVESTIYA SFedU. ENGINEERING SCIENCES

View full text Add to dashboard Cite

show abstract

“…Since the number of convolvers is fewer than the number of layers in the proposed accelerator, programmable line buffers (shown in Figure 2) are used to generate the mask of the processed layer. We employ line buffers [14], but add derivations (shown in blue in Figure 2) to adjust the data path (selected by the central control using multiplexers).…”

Section: Programmable Line Buffermentioning

confidence: 99%

“…State-of-the-art works are focused on reducing the amount of resources through two major design strategies. Single Convolver Architectures (SCA) is one of such strategies, where convolution layer are sequentially processed by adapting resources to the different layers hyper-parameters [13,14,15,16]. Multiple Convolver Architectures (MCA) are the second strategy that employs one convolver for each CNN convolutional layers, where a higher performance is achieved when multiple images are processed, but at the expense of a resources increment [17,18].…”

Section: Introductionmentioning

confidence: 99%

A streaming accelerator of Convolutional Neural Networks for resource-limited applications

Arredondo-Velázquez

Díaz-Carmona

Torres-Huitzil

et al. 2019

IEICE Electron. Express

View full text Add to dashboard Cite

Convolutional Neuronal Networks (CNN) implementation on embedded devices is restricted due to the number of layers of some CNN models. In this context, this paper describes a novel architecture based on Layer Operation Chaining (LOC) which uses fewer convolvers than convolution layers. A reutilization of hardware convolvers is promoted through kernel decomposition. Thus, an architectural design with reduced resources utilization is achieved, suitable to be implemented on low-end devices as a solution for portable classification applications. Experimental results show that the proposed design has a competitive processing time and overcomes resource utilization when compared with state-of-the-art related works.

show abstract

“…The feature extractor is used to filter input images into feature maps which are represented a set of features from the images. These features are represented a lowdimensional vector and include corners, lines, edges, etc., which are relatively invariant to position shifting or distortions [CS17]. Then the output from the feature extractor is fed into the classifier, which is usually based on traditional artificial neural networks.…”

Section: Convolutional Neural Networkmentioning

confidence: 99%

Human Action Recognition based on 3D Convolution Neural Networks from RGBD Videos

Al-Akam¹,

Paulus²,

Gharabaghi³

2018

CSRN

View full text Add to dashboard Cite

Human action recognition with color and depth sensors has received increasing attention in image processing and computer vision. This paper target is to develop a novel deep model for recognizing human action from the fusion of RGB-D videos based on a Convolutional Neural Network. This work is proposed a novel 3D Convolutional Neural Network architecture that implicitly captures motion information between adjacent frames, which are represented in two main steps: As a First, the optical flow is used to extract motion information from spatio-temporal domains of the different RGB-D video actions. This information is used to compute the features vector values from deep 3D CNN model. Secondly, train and evaluate a 3D CNN from three channels of the input video sequences (i.e. RGB, depth and combining information from both channels (RGB-D)) to obtain a feature representation for a 3D CNN model. For evaluating the accuracy results, a Convolutional Neural Network based on different data channels are trained and additionally the possibilities of feature extraction from 3D Convolutional Neural Network and the features are examined by support vector machine to improve and recognize human actions. From this methods, we demonstrate that the test results from RGB-D channels better than the results from each channel trained separately by baseline Convolutional Neural Network and outperform the state of the art on the same public datasets.

show abstract

An efficient implementation of 2D convolution in CNN

Cited by 34 publications

References 10 publications

Using Fast Prototyping Facilities for Implementation of a Convolution Neural Network on a Fpga

Using Fast Prototyping Facilities for Implementation of a Convolution Neural Network on a Fpga

A streaming accelerator of Convolutional Neural Networks for resource-limited applications

Human Action Recognition based on 3D Convolution Neural Networks from RGBD Videos

Contact Info

Product

Resources

About