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

Design method of multi-class center classification network model

A classification network and design method technology, applied in the field of deep learning, can solve problems such as model performance degradation

Active Publication Date: 2020-06-05
浙江芯劢微电子股份有限公司
View PDF4 Cites 3 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0009] The purpose of the present invention is to provide a design method of a classification network model of multi-class centers, by referring to the idea of ​​the mixed Gaussian model, a method of setting multiple cluster centers for each classification is proposed, and the optimal value selection is adopted Mechanism, so that different pictures with large feature differences in the same category can be gathered towards a most suitable feature center point for it, improve the speed of training, retain the diversity features of each category, and improve the model The problem of slow 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
  • Design method of multi-class center classification network model
  • Design method of multi-class center classification network model
  • Design method of multi-class center classification network model

Examples

Experimental program
Comparison scheme
Effect test

specific Embodiment approach 1

[0041] A design method of a classification network model of a multi-class center, comprising the steps of:

[0042] S1: Design a multi-class center classification network model, the model framework is as attached figure 2 As shown, the model framework of the classification network model includes a feature extraction network (CNN), a fully connected group, a multiplexer, and a loss function:

[0043] S11: Set multiple class centers for each category; after the CNN feature extraction network, connect a multi-channel parallel FC group, and the row vector of the weight matrix of each FC layer is a class center, that is, each category Corresponding to multiple row vectors in the multi-channel FC layer, there are multiple class centers; commonly used feature extraction networks include: Resnet, Inception, DenseNet, etc., and there are many versions of each type, which can be based on The scale of the task to be processed selects the feature extraction network, and selects the numb...

specific Embodiment approach 2

[0047] Optimize the penalty term of the loss function, and add the class center distance as an additional penalty term to the calculation of the loss function;

[0048] S21: During the forward calculation process, the aggregation radius of each center point in the dynamic statistics class, as shown in the attached Figure 4 As shown, in this example, the cosine distance is used as the measure of distance;

[0049] S22: Taking the aggregation radius as the threshold, when the distance between the centers in the class is smaller than the aggregation radius, add it as a penalty item to the loss function, which can drive the separation of the center points in the class and prevent them from sticking together. The trend of richness of features also increases the separability and determinability of the model. The calculation formula of the penalty item is:

[0050]

[0051]

[0052]

[0053]

[0054] Among them, N represents the number of target categories, L i.g12 , ...

specific Embodiment approach 3

[0057] Train the network model:

[0058] The forward calculation process is:

[0059] S211: Preprocessing the input samples, such as: image enhancement and normalization processing;

[0060] S212: Input the preprocessed image data into the feature network model, and calculate the corresponding feature vector;

[0061] S213: Send the feature vector to the parallel multi-channel FC layer and the row vector of the multi-channel FC to calculate the inner product value; as attached image 3 The operation of the eigenvector and the FC weight matrix shown, each row vector of each FC weight matrix is ​​used as a class center of a classification;

[0062] S214: Send the calculated value of the vector inner product in step S213 to the multiplexer to select a suitable class center; as attached image 3 As shown, the output value of the inner product of the feature vector and the FC row vector is the predicted probability of the category to which it belongs, and the specific appended ...

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 discloses a design method of a multi-class center classification network model, and relates to the technical field of deep learning. According to the method, the problem that the model performance is reduced when a CNN + FC + Softmax framework processes classification data sets with high feature diversity is solved; the thought of a Gaussian mixture model is used for reference; the invention provides a method for setting a plurality of clustering centers for each classification. An optimal value selection mechanism is adopted; different pictures with large feature differences inthe same class can be gathered towards a feature center point most suitable for the pictures; by means of the design method, all the center points in the class can be driven to be separated and prevented from being bonded together, the richness trend of class characteristics is increased, meanwhile, the separability and the judgeability of the model are improved, the training speed is increased, the diversity characteristics of all the classes are reserved, and the problem that the performance of the model is reduced is solved.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and in particular relates to a design method of a classification network model of a multi-class center. Background technique [0002] With the rapid development of deep learning technology, related technologies have been widely used in many fields. Among them, in the field of machine vision, it mainly includes: target detection, target classification, image segmentation and other key applications. The task of target classification is to confirm the category of the target in the image, which is one of the core issues in the field of machine vision and an important part of understanding image content in machine vision. [0003] The usual target classification tasks use the processing framework of CNN+FC (fully connected, fully connected)+Softmax. as attached figure 1 As shown in , it mainly includes: input image, feature extraction (CNN), full connection (FC), and loss function (SoftMax). ...

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): G06K9/62G06N3/04G06N3/08
CPCG06N3/084G06N3/045G06F18/23213G06F18/241
Inventor 葛益军
Owner 浙江芯劢微电子股份有限公司
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