A method for auxiliary data selection for small sample target recognition based on inverse transfer learning
The reverse transfer learning method enhances small sample target recognition by training on public datasets and selecting auxiliary data based on increased recognition rates, addressing the limitations of traditional methods and improving performance.
Patent Information
- Application Number
- CN202011442380.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-08
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2040-12-08
AI Technical Summary
Existing methods for selecting auxiliary data in small sample target recognition are complex and prone to negative transfer, limiting the scope of relevant data and the effectiveness of transfer learning.
Utilize a reverse transfer learning approach by training a backbone recognition network on a large-scale public dataset and then adding a small sample target dataset to enhance the network, selecting auxiliary data based on increased recognition rates in the public dataset.
Improves recognition performance by 4% in remote sensing ship target recognition tasks, overcoming limitations of traditional small sample learning methods.
Smart Images
Figure CN112541539B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of deep learning and relates to a method for selecting auxiliary data for small sample target recognition based on inverse transfer learning. Background Art
[0002] At present, target recognition has been widely used in military and civilian fields. In the past decade, the position of machine learning methods in target recognition has been gradually rising. Deep learning, as an important branch field, has more powerful feature learning and feature expression capabilities than traditional machine learning methods, and has better recognition effects.
[0003] However, in target recognition and even the entire deep learning direction, the classification ability of the model needs to be supported by massive amounts of data, but obtaining massive amounts of labeled data often consumes high manpower and time costs. Therefore, in target recognition tasks, we increasingly hope to use existing public datasets or other few-labeled or unlabeled data as auxiliary data when training deep network models, from which we can transfer some useful information for recognition tasks, so that we can train a better network model with only a small amount of labeled target data. Therefore, small sample-based target recognition methods have become a breakthrough in solving this problem.
[0004] The small sample learning method aims to extract additional feature information useful for solving the task from other related data, such as transfer learning. However, the existing methods are mainly transfer learning from candidate data sets to target data sets. Generally, the auxiliary data sets for the small sample target recognition task are selected by measuring the similarity between the candidate data sets and the target data sets. This kind of artificially defined explicit calculation is not only complex, but also has many limitations. If the inappropriate additional feature information is extracted and added to the small sample learning, it may have an adverse effect on the small sample learning task, such as "negative transfer" in transfer learning. Therefore, in small sample learning, how to more efficiently and accurately select auxiliary data to add to the small sample target recognition task is a very meaningful and urgent problem to be solved. Summary of the invention
[0005] In view of the deficiencies in the prior art, the present invention provides a method for selecting auxiliary data for small sample target recognition based on inverse transfer learning, and strives to break through the limitations of the existing small sample learning, that is, the scope of relevant data is limited and the migration effect is difficult to guarantee. This method is different from the traditional transfer learning from the candidate data set to the target data set. Instead, the target data set is added to the candidate data set for inverse transfer, and the auxiliary data is selected by using the deep network's own powerful data representation ability. Mainly, before the small sample recognition task, a large-scale and easily accessible public data set is used to train the backbone recognition network, and then the small sample target data set is added to the public data set to jointly train the backbone recognition network to achieve inverse transfer, so as to select the categories with more obvious recognition rate increase in the public data set as the auxiliary data for the small sample target recognition task, thereby realizing the selection of auxiliary data for the small sample target recognition task.
[0006] Step (1): Use the candidate data set to train the network.
[0007] Download existing and publicly available image datasets to train the backbone recognition network, test and record the detailed recognition results for each category in the image dataset;
[0008] Step (2): Reverse transfer learning
[0009] Add the small sample target data set to the image data set selected in step (1) to construct a new data set, use the new data set to jointly train the backbone recognition network to achieve reverse transfer learning, test and record the detailed recognition results of the backbone recognition network for the image data set selected in step (1) at this time;
[0010] Step (3), select auxiliary data set.
[0011] Compare the recognition results of the backbone recognition network on the image dataset selected in step (1) before and after reverse transfer learning, select the categories with more obvious recognition rate increase in the image dataset selected in step (1) as auxiliary data for the small sample target recognition task, and complete the selection of auxiliary data for the small sample target recognition task.
[0012] Preferably, the image dataset is one of ImageNet, miniImageNet, and PASCAL VOC.
[0013] The beneficial effects of the present invention are as follows: The key to the present invention is to break away from the traditional thinking of transfer learning from candidate data sets to target data sets. The target data set is added to the candidate data set, and the deep network's own powerful data representation ability is fully utilized. Through the reverse transfer method, effective small sample target recognition auxiliary data is accurately selected, thereby improving network performance. Experimental verification shows that using miniImageNet as a candidate data set and Resnet18 network as the backbone recognition network, the method proposed in the present invention can increase the recognition rate by 4% on the remote sensing ship target recognition task, breaking through the limitations of the existing small sample learning, which is limited in the scope of relevant data and difficult to guarantee the migration effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0015] The present invention is further analyzed below in conjunction with specific embodiments.
[0016] This experiment divides a set of collected remote sensing ship target images into training sets and test sets, uses miniImageNet as the candidate data set, and uses the Resnet18 network as the backbone recognition network. Figure 1 As shown in the figure, the specific steps in the remote sensing ship target recognition task are as follows:
[0017] Step (1): Use the candidate data set to train the network.
[0018] Use miniImageNet as a candidate dataset to train the Resnet18 network, test it, and record the detailed recognition results for each category.
[0019] Step (2): Reverse transfer learning
[0020] The remote sensing ship target dataset is added to the miniImageNet dataset to construct a new dataset. The new dataset is used to jointly train the Resnet18 network to achieve reverse transfer learning. The detailed recognition results of the recognition network for the miniImageNet dataset are tested and recorded.
[0021] Step (3), select auxiliary data set.
[0022] By comparing the recognition results of the Resnet18 network on the miniImageNet dataset before and after reverse migration, this experiment selected a total of 10 types of image data as auxiliary data for the remote sensing ship target recognition task.
[0023] After adding the auxiliary data selected in step (3) to the remote sensing ship dataset to train the backbone recognition network, comparative tests found that this method can effectively improve the recognition performance of the backbone recognition network and increase the recognition rate of the remote sensing ship target recognition task by 4%, breaking through the limitations of existing small sample learning, such as the limited range of relevant data and the difficulty in ensuring the transfer effect.
[0024] The above embodiments are not limitations of the present invention, and the present invention is not limited to the above embodiments. As long as the requirements of the present invention are met, they belong to the protection scope of the present invention.
Claims
1. A method for selecting auxiliary data for small sample target recognition based on inverse transfer learning. It is characterized in that The method specifically comprises the following steps: Step (1): Train the network using the candidate dataset Download an existing and publicly available image dataset to train the backbone recognition network, test and record the detailed recognition results of each category in the image dataset; the image dataset is one of ImageNet, miniImageNet, and PASCALVOC; the backbone recognition network is Resnet18; Step (2): Reverse transfer learning The remote sensing ship target dataset is added as a small sample target dataset to the image dataset selected in step (1) to construct a new dataset, and the backbone recognition network is jointly trained with the new dataset to implement reverse transfer learning. The detailed recognition results of the backbone recognition network for the image dataset selected in step (1) are tested and recorded at this time; Step (3): Select auxiliary dataset Compare the recognition results of the backbone recognition network on the image dataset selected in step (1) before and after reverse transfer learning, select the categories with more obvious recognition rate increase in the image dataset selected in step (1) as auxiliary data for the small sample target recognition task, and complete the selection of auxiliary data for the small sample target recognition task.