Training method and device of deep neural network based on hash coding layer

A deep neural network and hash coding technology, applied in the field of training methods and devices of deep neural networks, can solve the problems of difficulty in achieving optimal performance of coding and limited degree of non-linearity, etc.

Active Publication Date: 2017-12-15
NETPOSA TECH
View PDF2 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0004] However, in the above encoding methods, the degree of nonlinearization is limited, and since the binary model to be solved directly is a non-deterministic polynomial, it is generally necessary to relax the optimization conditions, and the encoding after conditional relaxation is difficult to achieve optimal performance

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
  • Training method and device of deep neural network based on hash coding layer
  • Training method and device of deep neural network based on hash coding layer
  • Training method and device of deep neural network based on hash coding layer

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0058] see Figure 1A , the embodiment of the present invention provides a method for training a deep neural network based on a hash coding layer. The method specifically includes the following steps:

[0059] Step 101: Obtain a pre-trained neural network model, which includes a basic feature layer;

[0060] The above-mentioned pre-trained neural network model can be an image classifier model trained based on a large-scale data set, or an image classification model obtained by using a large-scale data set to train a custom network structure through a softmax optimization function.

[0061] In the above-mentioned pre-trained neural network model, multiple layers from the first layer to the layer capable of extracting better image features are used as the basic feature layers. Before the method provided by the embodiment of the present invention performs the training based on the hash coding layer on the above-mentioned pre-trained neural network model, the basic feature layer ...

Embodiment 2

[0097] see figure 2 An embodiment of the present invention provides a training device for a deep neural network based on a hash coding layer, and the device is used to implement the training method for a deep neural network based on a hash coding layer provided in Embodiment 1 above. Specifically, the device includes:

[0098] An acquisition module 201, configured to acquire a pre-trained neural network model, the neural network model including a basic feature layer;

[0099] The first insertion module 202 is used to insert the first fully connected layer used for the hash code linear map and the hash code layer used as the activation function after the basic feature layer in the neural network model; use the hash code layer as The activation function of the deep neural network model stimulates the deep neural network through the hash coding layer to improve the accuracy of the deep neural network model.

[0100] The second insertion module 203 is used to insert the second ...

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 method and device for training a deep neural network based on a hash coding layer. Wherein, the method includes: obtaining a pre-trained neural network model, the neural network model includes a basic feature layer; after the basic feature layer in the neural network model, inserting the first fully connected layer used for hash coding linear mapping and used as an activation function; insert a second fully-connected layer after the hash-encoded layer as an output layer; train the weights of the first fully-connected layer and the second fully-connected layer. The present invention uses the hash coding layer as the activation function of the deep neural network training, and trains the neural network through the stochastic gradient descent method to achieve the overall optimization, and directly trains the output binary coded model without relaxing the optimization conditions change.

Description

technical field [0001] The invention relates to the fields of computer technology and image processing, in particular to a training method and device for a deep neural network based on a hash coding layer. Background technique [0002] In the field of image retrieval and image storage, since a large amount of data storage and feature comparison are involved, it is necessary to perform binary coding on the image on the basis of maintaining the similarity of the original image feature space to reduce the time spent on image retrieval and image storage. The amount of data to be processed. [0003] At present, in related technologies, images are binarized by hash coding technology, and hash coding technology can be divided into three categories according to the binarization coding method: 1) linear mapping coding; 2) semi-nonlinear coding; 3 ) nonlinear coding. Among them, the linear mapping coding method can be divided into three categories according to the type of data requi...

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 Patents(China)
IPC IPC(8): G06N3/08
Inventor 任鹏远许健万定锐
Owner NETPOSA TECH
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

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

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap