Feature Extraction Model Isomorphic Branch Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep, wide, or complex neural networks used for image classification and object detection often result in a surge of parameters and increased computation, making them unsuitable for portable devices and real-time applications due to storage and deployment challenges, as well as high real-time requirement scenarios where computation delays can be detrimental.
Innovation Solution
A method involving deriving homogeneous branches from a main network to create a homogeneous auxiliary training model, calculating knowledge synergy loss values based on inter-class distances, and adjusting parameters until convergence, then removing auxiliary branches to obtain a feature extraction model with only a main branch, which improves identification performance without increasing parameter or computation amounts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deeper, wider or more complex network structure is constructed to improve picture identification performance, then identification accuracy is improved, but the number of parameters surges making storage and deployment difficult for portable devices
Solution Approach 1:
The network is divided into a main network and multiple auxiliary training networks. Each auxiliary training network focuses on specific training tasks while sharing the main network backbone, thereby improving identification accuracy without proportionally increasing the total number of parameters.
Solution Approach 2:
The main network serves multiple functions by being shared across different auxiliary training networks. This allows the system to achieve complex identification tasks with a compact parameter set, as the same parameters serve multiple purposes in different training contexts.
2Measurement precision
If a deeper, wider or more complex network structure is constructed to improve picture identification performance, then identification accuracy is improved, but computation amount increases making real-time application difficult
Solution Approach 1:
The computation workload is segmented between the main network and auxiliary training networks. During inference, only the compact main network is deployed, achieving real-time performance. The auxiliary networks are used exclusively during the training phase to improve accuracy without impacting deployment complexity.
Solution Approach 2:
The auxiliary training networks perform preliminary training actions during the training phase to enhance feature extraction capabilities. This preliminary action improves identification accuracy without requiring the deployed model to have complex structures, enabling real-time inference with reduced computation.
3Measurement precision
If homogeneous auxiliary training networks are derived and used for training, then feature extraction performance is improved, but training model complexity increases temporarily
Solution Approach 1:
The auxiliary training networks are temporarily introduced during the training phase to improve feature extraction performance, then discarded after training completion. The main network retains the learned improvements without the auxiliary structures, achieving better performance with reduced complexity in the deployed model.
Data Source
AI summary
The present application discloses an image recognition method and apparatus, and a device and a readable storage medium. The method includes: obtaining a target image to be recognized; inputting the target image into a trained feature extraction model for feature extraction to obtain image features; and using the image features to recognize the target image to obtain a recognition result. In the present application, under the condition that the parameter quantity and the calculation amount of the feature extraction model do not need to be increased, the feature extraction model has relatively good feature extraction performance by means of isomorphic branch extraction and feature mutual mining, namely knowledge collaboration assistive training, and more accurate image recognition may be completed on the basis of the image features extracted by the feature extraction model.