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

Design method of image classification loss function based on cosine space optimization

A technology of loss function and design method, applied in the field of computer vision and artificial intelligence, can solve problems such as large parameter quantity and computational complexity, and achieve the effect of improving model performance

Pending Publication Date: 2021-06-29
SOUTHEAST UNIV
View PDF0 Cites 3 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0002] The target classification algorithm is a basic and important research field in computer vision. With the development of deep learning technology, the current target classification algorithm can solve most of the simple classification problems. After AlexNet, a series of CNN models such as ResNet have emerged. , GoogleNet, EfficientNet, constantly refreshing the results on the ImageNet dataset. With the adoption of more complex network structures and the introduction of deep residual connections, the current optimal image classification algorithm has reached the Top-1 accuracy on the ImageNet dataset. 84.4%, however these algorithmic models usually have a huge amount of parameters and computational complexity

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 image classification loss function based on cosine space optimization
  • Design method of image classification loss function based on cosine space optimization
  • Design method of image classification loss function based on cosine space optimization

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0037] This implementation provides an image classification loss function based on cosine space optimization. The loss function consists of three parts, and the specific expression is as follows:

[0038] Loss = αLoss 1 +β·Truc(Loss 2 -ε)+γLoss cross-entropy

[0039] In the formula, α is the weighting coefficient of the inter-class loss function, β is the weighting coefficient of the intra-class loss function, and γ is the weighting coefficient of the cross-entropy loss function, and the ratio of γ>α>β is adopted; Loss 1 Expressed as a between-class loss function, Loss cross-entropy Expressed as a cross-entropy loss function, Loss 2 Expressed as an intra-class loss function;

[0040] The hyperparameter β is a piecewise function. When the batch of algorithm model training is smaller than the hyperparameter N, β takes a value of 0, and when the training batch exceeds N, then β takes a value greater than 0.

[0041] The value of the hyperparameter N is determined according ...

Embodiment 2

[0055] This embodiment provides a method for designing an image classification loss function based on cosine space optimization, including the following steps:

[0056] Step S1, obtaining a data set, setting hyperparameters, and initializing a deep learning model;

[0057] Specifically, in this embodiment, the above hyperparameters include: weighting coefficient α, weighting coefficient β, weighting coefficient γ, compaction coefficient ε, and batch number N of iterative training, and satisfy: γ>α>β.

[0058] Step S2, input the acquired data set into the initialized deep learning model, perform multi-batch iterative training on the deep learning model, and execute step S21-step S23 sequentially in each iterative batch;

[0059] Step S21, according to the feature vector obtained by the deep learning model in the forward propagation process, calculate the intra-class center c of each category object in the current iteration batch i , and cumulatively update the in-class center ...

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 an image classification loss function based on cosine space optimization, and provides a loss function capable of actively regulating and controlling the intra-class distance and the inter-class distance of image classification at the same time on the basis of common Additive Margin Softmax optimization. According to the method, an AM-Softmax loss function is adopted in the first half stage of model training, the inter-class distance is increased, an intra-class center capable of being dynamically adjusted along with the training batch is added in the second half stage of training, feature vectors of objects of the same class are further tightened, meanwhile, the cosine distance between the feature vectors of objects of different classes is increased, the model can be converged more quickly, similar categories can be fully distinguished, and the model performance can be further improved.

Description

technical field [0001] The invention relates to the fields of computer vision and artificial intelligence, in particular to a design method of an image classification loss function based on cosine space optimization. Background technique [0002] The target classification algorithm is a basic and important research field in computer vision. With the development of deep learning technology, the current target classification algorithm can solve most of the simple classification problems. After AlexNet, a series of CNN models such as ResNet have emerged. , GoogleNet, EfficientNet, constantly refreshing the results on the ImageNet dataset. With the adoption of more complex network structures and the introduction of deep residual connections, the current optimal image classification algorithm has reached the Top-1 accuracy on the ImageNet dataset. 84.4%, however, these algorithmic models usually have a huge amount of parameters and computational complexity. For edge mobile scena...

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
CPCG06N3/084G06N3/047G06N3/044G06F18/214Y02T10/40
Inventor 李晨许虞俊孙翔曹悦欣杜文娟
Owner SOUTHEAST UNIV
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