Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Processing method and device of 3D convolutional neural network on neural network processor

A convolutional neural network and neural network technology, applied in the processing field of 3D convolutional neural network on the neural network processor, can solve the problems that the 3D convolutional neural network cannot be applied, limit the application of the 3D convolutional neural network, etc.

Active Publication Date: 2020-11-24
HANGZHOU HIKVISION DIGITAL TECH
View PDF8 Cites 1 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

When designing neural network processors, no expansion is considered. Neural network processors often only support two-dimensional convolutional neural networks (also known as 2D convolutions), and do not support convolutional neural networks with three-dimensional convolution kernels (referred to as three-dimensional convolutional neural networks). network, also known as 3D convolution), which leads to the fact that the three-dimensional convolutional neural network cannot be applied to the neural network processor, which limits the application of the three-dimensional convolutional neural network

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Processing method and device of 3D convolutional neural network on neural network processor
  • Processing method and device of 3D convolutional neural network on neural network processor
  • Processing method and device of 3D convolutional neural network on neural network processor

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0156] Embodiment 1: The second pooling operation is performed on the L first pooling feature maps along the direction of the quantity dimension to obtain the K feature maps compressed again, wherein the second pooling operation is in the Sliding the window between the L first pooling feature maps and performing the operation of value selection according to the set value method. It should be noted that, here, the way of setting the value may be taking the average value or taking the maximum value, and this embodiment takes the maximum value as an example for illustration.

[0157] In this embodiment, the L first pooled feature maps may be grouped according to the parameters of the secondary pooling window in the time dimension to obtain K groups, and each group includes L3 first pooled feature maps. In an example, the grouping can be performed by means of a sliding window, and the number K of groups can be calculated by formula (2).

[0158] K=(L-L3+2*Pad_L3) / STRIDE_L3+1 (2) ...

Embodiment 2

[0167] Embodiment 2: The L first pooling feature maps are first spliced ​​along the W direction to obtain a pooling splicing feature map after splicing along the W direction, and then the pooling splicing feature map after splicing along the W direction Perform the third pooling operation to obtain K feature maps compressed again; the third pooling operation is to slide the pooled spliced ​​feature maps spliced ​​along the W direction into a sliding window and take values ​​according to the set value method operation, the W direction is the width dimension direction. It should be noted that, here, the way of setting the value may be taking the average value or taking the maximum value, and this embodiment takes the maximum value as an example for illustration.

[0168]In this embodiment, the L first pooled feature maps are spliced ​​along the width dimension to obtain pooled spliced ​​feature maps.

[0169] During implementation, splicing can be performed in the order of the ...

Embodiment 3

[0181] Embodiment 3: splicing the L first pooling feature maps along the H direction to obtain a pooling splicing feature map after splicing along the H direction, and then performing the pooling splicing feature map after splicing along the H direction The fourth pooling operation is to obtain K feature maps compressed again. The fourth pooling operation is to slide the window of the pooled splicing feature maps spliced ​​along the H direction and take values ​​according to the set value method Operation, the H direction is the height dimension direction. It should be noted that, here, the way of setting the value may be taking the average value or taking the maximum value, and this embodiment takes the maximum value as an example for illustration.

[0182] In this embodiment, the L first pooled feature maps are spliced ​​along the height dimension to obtain pooled spliced ​​feature maps.

[0183] During implementation, splicing can be performed in the order of the width dim...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention provides a processing method and device of a 3D convolutional neural network on a neural network processor. The method comprises the following steps of: splitting a graph sequence in a time dimension, performing first convolution operation on the split graph sequence and a first convolution kernel of a P3D convolution layer to obtain a plurality of first 2D feature graphs, dividing the first 2D feature graphs, and splicing the divided first 2D feature graphs to a channel dimension to obtain a plurality of 2D spliced graphs; and meanwhile, splicing the data of a second convolutionkernel of the P3D convolution layer in the time dimension to the channel dimension to obtain a 2D spliced convolution kernel, and performing a second convolution operation based on the 2D spliced graph and the 2D spliced convolution kernel. Therefore, the neural network processor realizes convolution processing supporting the 3D neural network. Meanwhile, a P3D pooling layer is subjected to pooling operation step conversion, first pooling operation and second pooling operation are carried out respectively, and pooling processing of the 3D convolutional neural network supported by the neural network processor is realized.

Description

technical field [0001] The present application relates to the field of image processing, and in particular to a method and device for processing 3D convolutional neural networks on neural network processors. Background technique [0002] Convolutional Neural Networks (CNN) is a type of Feedforward Neural Networks (Feedforward Neural Networks) that includes convolution calculations and has a deep structure, and is widely used in various fields such as image recognition, speech recognition, and natural language recognition. . [0003] In convolutional neural network applications, neural network processors have become the first choice for applications due to their faster processing speed, especially in some application scenarios that require high real-time performance, neural network processors are more inclined to use to implement a convolutional neural network. [0004] However, since the hardware structure of the neural network processor is fixed, it generally only support...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06N3/04G06N3/063
CPCG06N3/063G06N3/045
Inventor 黄斌
Owner HANGZHOU HIKVISION DIGITAL TECH
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com