Neural network training method, picture classification method and picture classification system
A neural network training and neural network technology, which is applied in the field of image classification methods, image classification systems, and neural network training methods, can solve problems such as large models and slow reasoning speeds
Pending Publication Date: 2022-06-24
广州天宸健康科技有限公司
View PDF0 Cites 0 Cited by
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
But at present, UDA can only migrate from one field to another, which belongs to one-to-one migration, and because the model is relatively large, the reasoning speed is slow
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 moreImage
Smart Image Click on the blue labels to locate them in the text.
Smart ImageViewing Examples
Examples
Experimental program
Comparison scheme
Effect test
Embodiment Construction
[0037] The domain described in this application generally refers to the concept in transfer learning technology, that is, transferring knowledge from one domain (ie, the source domain) to another domain (ie, the target domain). It also includes pictures in different scenes. For example, pictures taken when the light is strong is one field, and the light is weaker is another field.
[0038] This application combines knowledge distillation and unsupervised domain adaptation methods to propose a new neural network training method, which can not only compress model parameters, but also improve accuracy, and effectively solve the problem of domain drift.
[0039] see figure 1 as well as figure 2 , the neural network training method of this application includes the steps:
[0040] S1: Prepare the dataset: source domain pictures and target domain pictures;
[0041] S2: Feature extraction:
[0042] Input the source domain picture and the target domain picture into the first neura...
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
Login to View More Abstract
The invention discloses a neural network training method, a picture classification method and a picture classification system. The neural network training method comprises the following steps: S1, preparing a data set including a source domain picture and a target domain picture; s2, feature extraction: inputting the source domain picture and the target domain picture into a first neural network to obtain a source domain feature SF1 and a target domain feature TF1; inputting the source domain picture and the target domain picture into a second neural network to obtain a source domain feature SF2 and a target domain feature TF2; s3, TDA loss of the source domain in the first neural network and TDA loss of the target domain in the second neural network are calculated respectively, and TDA loss TDA1 of the first neural network and TDA2 of the second neural network are obtained; s4, calculating knowledge distillation loss of the source domain and the target domain; and S5, respectively updating the first neural network and the second neural network. According to the method, model parameters can be compressed, the accuracy can be improved, and the field drift problem is effectively solved.
Description
technical field [0001] The invention relates to the technical field of information retrieval, in particular to a neural network training method, a picture classification method and a picture classification system. Background technique [0002] Deep learning is a technology that learns from data, and is currently widely used in the industry. Image classification based on deep learning has a high accuracy rate. [0003] The reason why deep learning is powerful is that it can learn features from data, so as to mine new rules, but deep learning also has its limitations, it is only applicable to the similar features in the data of the training set, when the same When a deep learning network or model is applied in a new field, it needs to be retrained or adjusted. That is to say, when migrating to a new field or when a new image appears, the field that did not appear in the original training set cannot be recognized. [0004] To address the limitations of deep learning, the curr...
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
Login to View More IPC IPC(8): G06N3/04G06N3/08G06N20/00G06V10/44G06V10/764G06V10/82G06K9/62
CPCG06N3/084G06N20/00G06N3/045G06F18/241
Inventor 曾祥云朱姬渊
Owner 广州天宸健康科技有限公司